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 return false; /* Caller handle free/refcnt */ 448 449 return true; 450 } 451 452 static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp) 453 { 454 struct virtnet_info *vi = netdev_priv(dev); 455 struct receive_queue *rq = vi->rq; 456 struct bpf_prog *xdp_prog; 457 bool sent; 458 459 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this 460 * indicate XDP resources have been successfully allocated. 461 */ 462 xdp_prog = rcu_dereference(rq->xdp_prog); 463 if (!xdp_prog) 464 return -ENXIO; 465 466 sent = __virtnet_xdp_xmit(vi, xdp); 467 if (!sent) 468 return -ENOSPC; 469 return 0; 470 } 471 472 static unsigned int virtnet_get_headroom(struct virtnet_info *vi) 473 { 474 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0; 475 } 476 477 /* We copy the packet for XDP in the following cases: 478 * 479 * 1) Packet is scattered across multiple rx buffers. 480 * 2) Headroom space is insufficient. 481 * 482 * This is inefficient but it's a temporary condition that 483 * we hit right after XDP is enabled and until queue is refilled 484 * with large buffers with sufficient headroom - so it should affect 485 * at most queue size packets. 486 * Afterwards, the conditions to enable 487 * XDP should preclude the underlying device from sending packets 488 * across multiple buffers (num_buf > 1), and we make sure buffers 489 * have enough headroom. 490 */ 491 static struct page *xdp_linearize_page(struct receive_queue *rq, 492 u16 *num_buf, 493 struct page *p, 494 int offset, 495 int page_off, 496 unsigned int *len) 497 { 498 struct page *page = alloc_page(GFP_ATOMIC); 499 500 if (!page) 501 return NULL; 502 503 memcpy(page_address(page) + page_off, page_address(p) + offset, *len); 504 page_off += *len; 505 506 while (--*num_buf) { 507 unsigned int buflen; 508 void *buf; 509 int off; 510 511 buf = virtqueue_get_buf(rq->vq, &buflen); 512 if (unlikely(!buf)) 513 goto err_buf; 514 515 p = virt_to_head_page(buf); 516 off = buf - page_address(p); 517 518 /* guard against a misconfigured or uncooperative backend that 519 * is sending packet larger than the MTU. 520 */ 521 if ((page_off + buflen) > PAGE_SIZE) { 522 put_page(p); 523 goto err_buf; 524 } 525 526 memcpy(page_address(page) + page_off, 527 page_address(p) + off, buflen); 528 page_off += buflen; 529 put_page(p); 530 } 531 532 /* Headroom does not contribute to packet length */ 533 *len = page_off - VIRTIO_XDP_HEADROOM; 534 return page; 535 err_buf: 536 __free_pages(page, 0); 537 return NULL; 538 } 539 540 static struct sk_buff *receive_small(struct net_device *dev, 541 struct virtnet_info *vi, 542 struct receive_queue *rq, 543 void *buf, void *ctx, 544 unsigned int len, 545 bool *xdp_xmit) 546 { 547 struct sk_buff *skb; 548 struct bpf_prog *xdp_prog; 549 unsigned int xdp_headroom = (unsigned long)ctx; 550 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom; 551 unsigned int headroom = vi->hdr_len + header_offset; 552 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 553 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 554 struct page *page = virt_to_head_page(buf); 555 unsigned int delta = 0; 556 struct page *xdp_page; 557 bool sent; 558 int err; 559 560 len -= vi->hdr_len; 561 562 rcu_read_lock(); 563 xdp_prog = rcu_dereference(rq->xdp_prog); 564 if (xdp_prog) { 565 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset; 566 struct xdp_buff xdp; 567 void *orig_data; 568 u32 act; 569 570 if (unlikely(hdr->hdr.gso_type)) 571 goto err_xdp; 572 573 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) { 574 int offset = buf - page_address(page) + header_offset; 575 unsigned int tlen = len + vi->hdr_len; 576 u16 num_buf = 1; 577 578 xdp_headroom = virtnet_get_headroom(vi); 579 header_offset = VIRTNET_RX_PAD + xdp_headroom; 580 headroom = vi->hdr_len + header_offset; 581 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 582 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 583 xdp_page = xdp_linearize_page(rq, &num_buf, page, 584 offset, header_offset, 585 &tlen); 586 if (!xdp_page) 587 goto err_xdp; 588 589 buf = page_address(xdp_page); 590 put_page(page); 591 page = xdp_page; 592 } 593 594 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len; 595 xdp.data = xdp.data_hard_start + xdp_headroom; 596 xdp_set_data_meta_invalid(&xdp); 597 xdp.data_end = xdp.data + len; 598 xdp.rxq = &rq->xdp_rxq; 599 orig_data = xdp.data; 600 act = bpf_prog_run_xdp(xdp_prog, &xdp); 601 602 switch (act) { 603 case XDP_PASS: 604 /* Recalculate length in case bpf program changed it */ 605 delta = orig_data - xdp.data; 606 break; 607 case XDP_TX: 608 sent = __virtnet_xdp_xmit(vi, &xdp); 609 if (unlikely(!sent)) { 610 trace_xdp_exception(vi->dev, xdp_prog, act); 611 goto err_xdp; 612 } 613 *xdp_xmit = true; 614 rcu_read_unlock(); 615 goto xdp_xmit; 616 case XDP_REDIRECT: 617 err = xdp_do_redirect(dev, &xdp, xdp_prog); 618 if (err) 619 goto err_xdp; 620 *xdp_xmit = true; 621 rcu_read_unlock(); 622 goto xdp_xmit; 623 default: 624 bpf_warn_invalid_xdp_action(act); 625 case XDP_ABORTED: 626 trace_xdp_exception(vi->dev, xdp_prog, act); 627 case XDP_DROP: 628 goto err_xdp; 629 } 630 } 631 rcu_read_unlock(); 632 633 skb = build_skb(buf, buflen); 634 if (!skb) { 635 put_page(page); 636 goto err; 637 } 638 skb_reserve(skb, headroom - delta); 639 skb_put(skb, len + delta); 640 if (!delta) { 641 buf += header_offset; 642 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len); 643 } /* keep zeroed vnet hdr since packet was changed by bpf */ 644 645 err: 646 return skb; 647 648 err_xdp: 649 rcu_read_unlock(); 650 dev->stats.rx_dropped++; 651 put_page(page); 652 xdp_xmit: 653 return NULL; 654 } 655 656 static struct sk_buff *receive_big(struct net_device *dev, 657 struct virtnet_info *vi, 658 struct receive_queue *rq, 659 void *buf, 660 unsigned int len) 661 { 662 struct page *page = buf; 663 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE); 664 665 if (unlikely(!skb)) 666 goto err; 667 668 return skb; 669 670 err: 671 dev->stats.rx_dropped++; 672 give_pages(rq, page); 673 return NULL; 674 } 675 676 static struct sk_buff *receive_mergeable(struct net_device *dev, 677 struct virtnet_info *vi, 678 struct receive_queue *rq, 679 void *buf, 680 void *ctx, 681 unsigned int len, 682 bool *xdp_xmit) 683 { 684 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 685 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 686 struct page *page = virt_to_head_page(buf); 687 int offset = buf - page_address(page); 688 struct sk_buff *head_skb, *curr_skb; 689 struct bpf_prog *xdp_prog; 690 unsigned int truesize; 691 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 692 bool sent; 693 694 head_skb = NULL; 695 696 rcu_read_lock(); 697 xdp_prog = rcu_dereference(rq->xdp_prog); 698 if (xdp_prog) { 699 struct page *xdp_page; 700 struct xdp_buff xdp; 701 void *data; 702 u32 act; 703 704 /* This happens when rx buffer size is underestimated */ 705 if (unlikely(num_buf > 1 || 706 headroom < virtnet_get_headroom(vi))) { 707 /* linearize data for XDP */ 708 xdp_page = xdp_linearize_page(rq, &num_buf, 709 page, offset, 710 VIRTIO_XDP_HEADROOM, 711 &len); 712 if (!xdp_page) 713 goto err_xdp; 714 offset = VIRTIO_XDP_HEADROOM; 715 } else { 716 xdp_page = page; 717 } 718 719 /* Transient failure which in theory could occur if 720 * in-flight packets from before XDP was enabled reach 721 * the receive path after XDP is loaded. In practice I 722 * was not able to create this condition. 723 */ 724 if (unlikely(hdr->hdr.gso_type)) 725 goto err_xdp; 726 727 /* Allow consuming headroom but reserve enough space to push 728 * the descriptor on if we get an XDP_TX return code. 729 */ 730 data = page_address(xdp_page) + offset; 731 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len; 732 xdp.data = data + vi->hdr_len; 733 xdp_set_data_meta_invalid(&xdp); 734 xdp.data_end = xdp.data + (len - vi->hdr_len); 735 xdp.rxq = &rq->xdp_rxq; 736 737 act = bpf_prog_run_xdp(xdp_prog, &xdp); 738 739 if (act != XDP_PASS) 740 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len); 741 742 switch (act) { 743 case XDP_PASS: 744 /* recalculate offset to account for any header 745 * adjustments. Note other cases do not build an 746 * skb and avoid using offset 747 */ 748 offset = xdp.data - 749 page_address(xdp_page) - vi->hdr_len; 750 751 /* We can only create skb based on xdp_page. */ 752 if (unlikely(xdp_page != page)) { 753 rcu_read_unlock(); 754 put_page(page); 755 head_skb = page_to_skb(vi, rq, xdp_page, 756 offset, len, PAGE_SIZE); 757 return head_skb; 758 } 759 break; 760 case XDP_TX: 761 sent = __virtnet_xdp_xmit(vi, &xdp); 762 if (unlikely(!sent)) { 763 trace_xdp_exception(vi->dev, xdp_prog, act); 764 if (unlikely(xdp_page != page)) 765 put_page(xdp_page); 766 goto err_xdp; 767 } 768 *xdp_xmit = true; 769 if (unlikely(xdp_page != page)) 770 goto err_xdp; 771 rcu_read_unlock(); 772 goto xdp_xmit; 773 default: 774 bpf_warn_invalid_xdp_action(act); 775 case XDP_ABORTED: 776 trace_xdp_exception(vi->dev, xdp_prog, act); 777 case XDP_DROP: 778 if (unlikely(xdp_page != page)) 779 __free_pages(xdp_page, 0); 780 goto err_xdp; 781 } 782 } 783 rcu_read_unlock(); 784 785 truesize = mergeable_ctx_to_truesize(ctx); 786 if (unlikely(len > truesize)) { 787 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 788 dev->name, len, (unsigned long)ctx); 789 dev->stats.rx_length_errors++; 790 goto err_skb; 791 } 792 793 head_skb = page_to_skb(vi, rq, page, offset, len, truesize); 794 curr_skb = head_skb; 795 796 if (unlikely(!curr_skb)) 797 goto err_skb; 798 while (--num_buf) { 799 int num_skb_frags; 800 801 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 802 if (unlikely(!buf)) { 803 pr_debug("%s: rx error: %d buffers out of %d missing\n", 804 dev->name, num_buf, 805 virtio16_to_cpu(vi->vdev, 806 hdr->num_buffers)); 807 dev->stats.rx_length_errors++; 808 goto err_buf; 809 } 810 811 page = virt_to_head_page(buf); 812 813 truesize = mergeable_ctx_to_truesize(ctx); 814 if (unlikely(len > truesize)) { 815 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 816 dev->name, len, (unsigned long)ctx); 817 dev->stats.rx_length_errors++; 818 goto err_skb; 819 } 820 821 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 822 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 823 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 824 825 if (unlikely(!nskb)) 826 goto err_skb; 827 if (curr_skb == head_skb) 828 skb_shinfo(curr_skb)->frag_list = nskb; 829 else 830 curr_skb->next = nskb; 831 curr_skb = nskb; 832 head_skb->truesize += nskb->truesize; 833 num_skb_frags = 0; 834 } 835 if (curr_skb != head_skb) { 836 head_skb->data_len += len; 837 head_skb->len += len; 838 head_skb->truesize += truesize; 839 } 840 offset = buf - page_address(page); 841 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 842 put_page(page); 843 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 844 len, truesize); 845 } else { 846 skb_add_rx_frag(curr_skb, num_skb_frags, page, 847 offset, len, truesize); 848 } 849 } 850 851 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 852 return head_skb; 853 854 err_xdp: 855 rcu_read_unlock(); 856 err_skb: 857 put_page(page); 858 while (--num_buf) { 859 buf = virtqueue_get_buf(rq->vq, &len); 860 if (unlikely(!buf)) { 861 pr_debug("%s: rx error: %d buffers missing\n", 862 dev->name, num_buf); 863 dev->stats.rx_length_errors++; 864 break; 865 } 866 page = virt_to_head_page(buf); 867 put_page(page); 868 } 869 err_buf: 870 dev->stats.rx_dropped++; 871 dev_kfree_skb(head_skb); 872 xdp_xmit: 873 return NULL; 874 } 875 876 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 877 void *buf, unsigned int len, void **ctx, bool *xdp_xmit) 878 { 879 struct net_device *dev = vi->dev; 880 struct sk_buff *skb; 881 struct virtio_net_hdr_mrg_rxbuf *hdr; 882 int ret; 883 884 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 885 pr_debug("%s: short packet %i\n", dev->name, len); 886 dev->stats.rx_length_errors++; 887 if (vi->mergeable_rx_bufs) { 888 put_page(virt_to_head_page(buf)); 889 } else if (vi->big_packets) { 890 give_pages(rq, buf); 891 } else { 892 put_page(virt_to_head_page(buf)); 893 } 894 return 0; 895 } 896 897 if (vi->mergeable_rx_bufs) 898 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit); 899 else if (vi->big_packets) 900 skb = receive_big(dev, vi, rq, buf, len); 901 else 902 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit); 903 904 if (unlikely(!skb)) 905 return 0; 906 907 hdr = skb_vnet_hdr(skb); 908 909 ret = skb->len; 910 911 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 912 skb->ip_summed = CHECKSUM_UNNECESSARY; 913 914 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 915 virtio_is_little_endian(vi->vdev))) { 916 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 917 dev->name, hdr->hdr.gso_type, 918 hdr->hdr.gso_size); 919 goto frame_err; 920 } 921 922 skb->protocol = eth_type_trans(skb, dev); 923 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 924 ntohs(skb->protocol), skb->len, skb->pkt_type); 925 926 napi_gro_receive(&rq->napi, skb); 927 return ret; 928 929 frame_err: 930 dev->stats.rx_frame_errors++; 931 dev_kfree_skb(skb); 932 return 0; 933 } 934 935 /* Unlike mergeable buffers, all buffers are allocated to the 936 * same size, except for the headroom. For this reason we do 937 * not need to use mergeable_len_to_ctx here - it is enough 938 * to store the headroom as the context ignoring the truesize. 939 */ 940 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 941 gfp_t gfp) 942 { 943 struct page_frag *alloc_frag = &rq->alloc_frag; 944 char *buf; 945 unsigned int xdp_headroom = virtnet_get_headroom(vi); 946 void *ctx = (void *)(unsigned long)xdp_headroom; 947 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 948 int err; 949 950 len = SKB_DATA_ALIGN(len) + 951 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 952 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 953 return -ENOMEM; 954 955 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 956 get_page(alloc_frag->page); 957 alloc_frag->offset += len; 958 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 959 vi->hdr_len + GOOD_PACKET_LEN); 960 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 961 if (err < 0) 962 put_page(virt_to_head_page(buf)); 963 return err; 964 } 965 966 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 967 gfp_t gfp) 968 { 969 struct page *first, *list = NULL; 970 char *p; 971 int i, err, offset; 972 973 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2); 974 975 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */ 976 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) { 977 first = get_a_page(rq, gfp); 978 if (!first) { 979 if (list) 980 give_pages(rq, list); 981 return -ENOMEM; 982 } 983 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 984 985 /* chain new page in list head to match sg */ 986 first->private = (unsigned long)list; 987 list = first; 988 } 989 990 first = get_a_page(rq, gfp); 991 if (!first) { 992 give_pages(rq, list); 993 return -ENOMEM; 994 } 995 p = page_address(first); 996 997 /* rq->sg[0], rq->sg[1] share the same page */ 998 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 999 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 1000 1001 /* rq->sg[1] for data packet, from offset */ 1002 offset = sizeof(struct padded_vnet_hdr); 1003 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 1004 1005 /* chain first in list head */ 1006 first->private = (unsigned long)list; 1007 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2, 1008 first, gfp); 1009 if (err < 0) 1010 give_pages(rq, first); 1011 1012 return err; 1013 } 1014 1015 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 1016 struct ewma_pkt_len *avg_pkt_len) 1017 { 1018 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 1019 unsigned int len; 1020 1021 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1022 rq->min_buf_len, PAGE_SIZE - hdr_len); 1023 return ALIGN(len, L1_CACHE_BYTES); 1024 } 1025 1026 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1027 struct receive_queue *rq, gfp_t gfp) 1028 { 1029 struct page_frag *alloc_frag = &rq->alloc_frag; 1030 unsigned int headroom = virtnet_get_headroom(vi); 1031 char *buf; 1032 void *ctx; 1033 int err; 1034 unsigned int len, hole; 1035 1036 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len); 1037 if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp))) 1038 return -ENOMEM; 1039 1040 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1041 buf += headroom; /* advance address leaving hole at front of pkt */ 1042 get_page(alloc_frag->page); 1043 alloc_frag->offset += len + headroom; 1044 hole = alloc_frag->size - alloc_frag->offset; 1045 if (hole < len + headroom) { 1046 /* To avoid internal fragmentation, if there is very likely not 1047 * enough space for another buffer, add the remaining space to 1048 * the current buffer. 1049 */ 1050 len += hole; 1051 alloc_frag->offset += hole; 1052 } 1053 1054 sg_init_one(rq->sg, buf, len); 1055 ctx = mergeable_len_to_ctx(len, headroom); 1056 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1057 if (err < 0) 1058 put_page(virt_to_head_page(buf)); 1059 1060 return err; 1061 } 1062 1063 /* 1064 * Returns false if we couldn't fill entirely (OOM). 1065 * 1066 * Normally run in the receive path, but can also be run from ndo_open 1067 * before we're receiving packets, or from refill_work which is 1068 * careful to disable receiving (using napi_disable). 1069 */ 1070 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1071 gfp_t gfp) 1072 { 1073 int err; 1074 bool oom; 1075 1076 do { 1077 if (vi->mergeable_rx_bufs) 1078 err = add_recvbuf_mergeable(vi, rq, gfp); 1079 else if (vi->big_packets) 1080 err = add_recvbuf_big(vi, rq, gfp); 1081 else 1082 err = add_recvbuf_small(vi, rq, gfp); 1083 1084 oom = err == -ENOMEM; 1085 if (err) 1086 break; 1087 } while (rq->vq->num_free); 1088 virtqueue_kick(rq->vq); 1089 return !oom; 1090 } 1091 1092 static void skb_recv_done(struct virtqueue *rvq) 1093 { 1094 struct virtnet_info *vi = rvq->vdev->priv; 1095 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1096 1097 virtqueue_napi_schedule(&rq->napi, rvq); 1098 } 1099 1100 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1101 { 1102 napi_enable(napi); 1103 1104 /* If all buffers were filled by other side before we napi_enabled, we 1105 * won't get another interrupt, so process any outstanding packets now. 1106 * Call local_bh_enable after to trigger softIRQ processing. 1107 */ 1108 local_bh_disable(); 1109 virtqueue_napi_schedule(napi, vq); 1110 local_bh_enable(); 1111 } 1112 1113 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1114 struct virtqueue *vq, 1115 struct napi_struct *napi) 1116 { 1117 if (!napi->weight) 1118 return; 1119 1120 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1121 * enable the feature if this is likely affine with the transmit path. 1122 */ 1123 if (!vi->affinity_hint_set) { 1124 napi->weight = 0; 1125 return; 1126 } 1127 1128 return virtnet_napi_enable(vq, napi); 1129 } 1130 1131 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1132 { 1133 if (napi->weight) 1134 napi_disable(napi); 1135 } 1136 1137 static void refill_work(struct work_struct *work) 1138 { 1139 struct virtnet_info *vi = 1140 container_of(work, struct virtnet_info, refill.work); 1141 bool still_empty; 1142 int i; 1143 1144 for (i = 0; i < vi->curr_queue_pairs; i++) { 1145 struct receive_queue *rq = &vi->rq[i]; 1146 1147 napi_disable(&rq->napi); 1148 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1149 virtnet_napi_enable(rq->vq, &rq->napi); 1150 1151 /* In theory, this can happen: if we don't get any buffers in 1152 * we will *never* try to fill again. 1153 */ 1154 if (still_empty) 1155 schedule_delayed_work(&vi->refill, HZ/2); 1156 } 1157 } 1158 1159 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit) 1160 { 1161 struct virtnet_info *vi = rq->vq->vdev->priv; 1162 unsigned int len, received = 0, bytes = 0; 1163 void *buf; 1164 1165 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1166 void *ctx; 1167 1168 while (received < budget && 1169 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1170 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit); 1171 received++; 1172 } 1173 } else { 1174 while (received < budget && 1175 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1176 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit); 1177 received++; 1178 } 1179 } 1180 1181 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) { 1182 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) 1183 schedule_delayed_work(&vi->refill, 0); 1184 } 1185 1186 u64_stats_update_begin(&rq->stats.syncp); 1187 rq->stats.bytes += bytes; 1188 rq->stats.packets += received; 1189 u64_stats_update_end(&rq->stats.syncp); 1190 1191 return received; 1192 } 1193 1194 static void free_old_xmit_skbs(struct send_queue *sq) 1195 { 1196 struct sk_buff *skb; 1197 unsigned int len; 1198 unsigned int packets = 0; 1199 unsigned int bytes = 0; 1200 1201 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) { 1202 pr_debug("Sent skb %p\n", skb); 1203 1204 bytes += skb->len; 1205 packets++; 1206 1207 dev_consume_skb_any(skb); 1208 } 1209 1210 /* Avoid overhead when no packets have been processed 1211 * happens when called speculatively from start_xmit. 1212 */ 1213 if (!packets) 1214 return; 1215 1216 u64_stats_update_begin(&sq->stats.syncp); 1217 sq->stats.bytes += bytes; 1218 sq->stats.packets += packets; 1219 u64_stats_update_end(&sq->stats.syncp); 1220 } 1221 1222 static void virtnet_poll_cleantx(struct receive_queue *rq) 1223 { 1224 struct virtnet_info *vi = rq->vq->vdev->priv; 1225 unsigned int index = vq2rxq(rq->vq); 1226 struct send_queue *sq = &vi->sq[index]; 1227 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1228 1229 if (!sq->napi.weight) 1230 return; 1231 1232 if (__netif_tx_trylock(txq)) { 1233 free_old_xmit_skbs(sq); 1234 __netif_tx_unlock(txq); 1235 } 1236 1237 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1238 netif_tx_wake_queue(txq); 1239 } 1240 1241 static int virtnet_poll(struct napi_struct *napi, int budget) 1242 { 1243 struct receive_queue *rq = 1244 container_of(napi, struct receive_queue, napi); 1245 unsigned int received; 1246 bool xdp_xmit = false; 1247 1248 virtnet_poll_cleantx(rq); 1249 1250 received = virtnet_receive(rq, budget, &xdp_xmit); 1251 1252 /* Out of packets? */ 1253 if (received < budget) 1254 virtqueue_napi_complete(napi, rq->vq, received); 1255 1256 if (xdp_xmit) 1257 xdp_do_flush_map(); 1258 1259 return received; 1260 } 1261 1262 static int virtnet_open(struct net_device *dev) 1263 { 1264 struct virtnet_info *vi = netdev_priv(dev); 1265 int i, err; 1266 1267 for (i = 0; i < vi->max_queue_pairs; i++) { 1268 if (i < vi->curr_queue_pairs) 1269 /* Make sure we have some buffers: if oom use wq. */ 1270 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1271 schedule_delayed_work(&vi->refill, 0); 1272 1273 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i); 1274 if (err < 0) 1275 return err; 1276 1277 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 1278 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 1279 } 1280 1281 return 0; 1282 } 1283 1284 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 1285 { 1286 struct send_queue *sq = container_of(napi, struct send_queue, napi); 1287 struct virtnet_info *vi = sq->vq->vdev->priv; 1288 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq)); 1289 1290 __netif_tx_lock(txq, raw_smp_processor_id()); 1291 free_old_xmit_skbs(sq); 1292 __netif_tx_unlock(txq); 1293 1294 virtqueue_napi_complete(napi, sq->vq, 0); 1295 1296 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1297 netif_tx_wake_queue(txq); 1298 1299 return 0; 1300 } 1301 1302 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 1303 { 1304 struct virtio_net_hdr_mrg_rxbuf *hdr; 1305 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 1306 struct virtnet_info *vi = sq->vq->vdev->priv; 1307 int num_sg; 1308 unsigned hdr_len = vi->hdr_len; 1309 bool can_push; 1310 1311 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 1312 1313 can_push = vi->any_header_sg && 1314 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 1315 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 1316 /* Even if we can, don't push here yet as this would skew 1317 * csum_start offset below. */ 1318 if (can_push) 1319 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 1320 else 1321 hdr = skb_vnet_hdr(skb); 1322 1323 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 1324 virtio_is_little_endian(vi->vdev), false)) 1325 BUG(); 1326 1327 if (vi->mergeable_rx_bufs) 1328 hdr->num_buffers = 0; 1329 1330 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 1331 if (can_push) { 1332 __skb_push(skb, hdr_len); 1333 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 1334 if (unlikely(num_sg < 0)) 1335 return num_sg; 1336 /* Pull header back to avoid skew in tx bytes calculations. */ 1337 __skb_pull(skb, hdr_len); 1338 } else { 1339 sg_set_buf(sq->sg, hdr, hdr_len); 1340 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 1341 if (unlikely(num_sg < 0)) 1342 return num_sg; 1343 num_sg++; 1344 } 1345 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 1346 } 1347 1348 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 1349 { 1350 struct virtnet_info *vi = netdev_priv(dev); 1351 int qnum = skb_get_queue_mapping(skb); 1352 struct send_queue *sq = &vi->sq[qnum]; 1353 int err; 1354 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 1355 bool kick = !skb->xmit_more; 1356 bool use_napi = sq->napi.weight; 1357 1358 /* Free up any pending old buffers before queueing new ones. */ 1359 free_old_xmit_skbs(sq); 1360 1361 if (use_napi && kick) 1362 virtqueue_enable_cb_delayed(sq->vq); 1363 1364 /* timestamp packet in software */ 1365 skb_tx_timestamp(skb); 1366 1367 /* Try to transmit */ 1368 err = xmit_skb(sq, skb); 1369 1370 /* This should not happen! */ 1371 if (unlikely(err)) { 1372 dev->stats.tx_fifo_errors++; 1373 if (net_ratelimit()) 1374 dev_warn(&dev->dev, 1375 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err); 1376 dev->stats.tx_dropped++; 1377 dev_kfree_skb_any(skb); 1378 return NETDEV_TX_OK; 1379 } 1380 1381 /* Don't wait up for transmitted skbs to be freed. */ 1382 if (!use_napi) { 1383 skb_orphan(skb); 1384 nf_reset(skb); 1385 } 1386 1387 /* If running out of space, stop queue to avoid getting packets that we 1388 * are then unable to transmit. 1389 * An alternative would be to force queuing layer to requeue the skb by 1390 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be 1391 * returned in a normal path of operation: it means that driver is not 1392 * maintaining the TX queue stop/start state properly, and causes 1393 * the stack to do a non-trivial amount of useless work. 1394 * Since most packets only take 1 or 2 ring slots, stopping the queue 1395 * early means 16 slots are typically wasted. 1396 */ 1397 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { 1398 netif_stop_subqueue(dev, qnum); 1399 if (!use_napi && 1400 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { 1401 /* More just got used, free them then recheck. */ 1402 free_old_xmit_skbs(sq); 1403 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { 1404 netif_start_subqueue(dev, qnum); 1405 virtqueue_disable_cb(sq->vq); 1406 } 1407 } 1408 } 1409 1410 if (kick || netif_xmit_stopped(txq)) 1411 virtqueue_kick(sq->vq); 1412 1413 return NETDEV_TX_OK; 1414 } 1415 1416 /* 1417 * Send command via the control virtqueue and check status. Commands 1418 * supported by the hypervisor, as indicated by feature bits, should 1419 * never fail unless improperly formatted. 1420 */ 1421 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 1422 struct scatterlist *out) 1423 { 1424 struct scatterlist *sgs[4], hdr, stat; 1425 unsigned out_num = 0, tmp; 1426 1427 /* Caller should know better */ 1428 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 1429 1430 vi->ctrl_status = ~0; 1431 vi->ctrl_hdr.class = class; 1432 vi->ctrl_hdr.cmd = cmd; 1433 /* Add header */ 1434 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr)); 1435 sgs[out_num++] = &hdr; 1436 1437 if (out) 1438 sgs[out_num++] = out; 1439 1440 /* Add return status. */ 1441 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status)); 1442 sgs[out_num] = &stat; 1443 1444 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 1445 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 1446 1447 if (unlikely(!virtqueue_kick(vi->cvq))) 1448 return vi->ctrl_status == VIRTIO_NET_OK; 1449 1450 /* Spin for a response, the kick causes an ioport write, trapping 1451 * into the hypervisor, so the request should be handled immediately. 1452 */ 1453 while (!virtqueue_get_buf(vi->cvq, &tmp) && 1454 !virtqueue_is_broken(vi->cvq)) 1455 cpu_relax(); 1456 1457 return vi->ctrl_status == VIRTIO_NET_OK; 1458 } 1459 1460 static int virtnet_set_mac_address(struct net_device *dev, void *p) 1461 { 1462 struct virtnet_info *vi = netdev_priv(dev); 1463 struct virtio_device *vdev = vi->vdev; 1464 int ret; 1465 struct sockaddr *addr; 1466 struct scatterlist sg; 1467 1468 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 1469 if (!addr) 1470 return -ENOMEM; 1471 1472 ret = eth_prepare_mac_addr_change(dev, addr); 1473 if (ret) 1474 goto out; 1475 1476 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 1477 sg_init_one(&sg, addr->sa_data, dev->addr_len); 1478 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1479 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 1480 dev_warn(&vdev->dev, 1481 "Failed to set mac address by vq command.\n"); 1482 ret = -EINVAL; 1483 goto out; 1484 } 1485 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 1486 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 1487 unsigned int i; 1488 1489 /* Naturally, this has an atomicity problem. */ 1490 for (i = 0; i < dev->addr_len; i++) 1491 virtio_cwrite8(vdev, 1492 offsetof(struct virtio_net_config, mac) + 1493 i, addr->sa_data[i]); 1494 } 1495 1496 eth_commit_mac_addr_change(dev, p); 1497 ret = 0; 1498 1499 out: 1500 kfree(addr); 1501 return ret; 1502 } 1503 1504 static void virtnet_stats(struct net_device *dev, 1505 struct rtnl_link_stats64 *tot) 1506 { 1507 struct virtnet_info *vi = netdev_priv(dev); 1508 unsigned int start; 1509 int i; 1510 1511 for (i = 0; i < vi->max_queue_pairs; i++) { 1512 u64 tpackets, tbytes, rpackets, rbytes; 1513 struct receive_queue *rq = &vi->rq[i]; 1514 struct send_queue *sq = &vi->sq[i]; 1515 1516 do { 1517 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1518 tpackets = sq->stats.packets; 1519 tbytes = sq->stats.bytes; 1520 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1521 1522 do { 1523 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1524 rpackets = rq->stats.packets; 1525 rbytes = rq->stats.bytes; 1526 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 1527 1528 tot->rx_packets += rpackets; 1529 tot->tx_packets += tpackets; 1530 tot->rx_bytes += rbytes; 1531 tot->tx_bytes += tbytes; 1532 } 1533 1534 tot->tx_dropped = dev->stats.tx_dropped; 1535 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 1536 tot->rx_dropped = dev->stats.rx_dropped; 1537 tot->rx_length_errors = dev->stats.rx_length_errors; 1538 tot->rx_frame_errors = dev->stats.rx_frame_errors; 1539 } 1540 1541 #ifdef CONFIG_NET_POLL_CONTROLLER 1542 static void virtnet_netpoll(struct net_device *dev) 1543 { 1544 struct virtnet_info *vi = netdev_priv(dev); 1545 int i; 1546 1547 for (i = 0; i < vi->curr_queue_pairs; i++) 1548 napi_schedule(&vi->rq[i].napi); 1549 } 1550 #endif 1551 1552 static void virtnet_ack_link_announce(struct virtnet_info *vi) 1553 { 1554 rtnl_lock(); 1555 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 1556 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 1557 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 1558 rtnl_unlock(); 1559 } 1560 1561 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1562 { 1563 struct scatterlist sg; 1564 struct net_device *dev = vi->dev; 1565 1566 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 1567 return 0; 1568 1569 vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 1570 sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq)); 1571 1572 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 1573 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 1574 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 1575 queue_pairs); 1576 return -EINVAL; 1577 } else { 1578 vi->curr_queue_pairs = queue_pairs; 1579 /* virtnet_open() will refill when device is going to up. */ 1580 if (dev->flags & IFF_UP) 1581 schedule_delayed_work(&vi->refill, 0); 1582 } 1583 1584 return 0; 1585 } 1586 1587 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1588 { 1589 int err; 1590 1591 rtnl_lock(); 1592 err = _virtnet_set_queues(vi, queue_pairs); 1593 rtnl_unlock(); 1594 return err; 1595 } 1596 1597 static int virtnet_close(struct net_device *dev) 1598 { 1599 struct virtnet_info *vi = netdev_priv(dev); 1600 int i; 1601 1602 /* Make sure refill_work doesn't re-enable napi! */ 1603 cancel_delayed_work_sync(&vi->refill); 1604 1605 for (i = 0; i < vi->max_queue_pairs; i++) { 1606 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1607 napi_disable(&vi->rq[i].napi); 1608 virtnet_napi_tx_disable(&vi->sq[i].napi); 1609 } 1610 1611 return 0; 1612 } 1613 1614 static void virtnet_set_rx_mode(struct net_device *dev) 1615 { 1616 struct virtnet_info *vi = netdev_priv(dev); 1617 struct scatterlist sg[2]; 1618 struct virtio_net_ctrl_mac *mac_data; 1619 struct netdev_hw_addr *ha; 1620 int uc_count; 1621 int mc_count; 1622 void *buf; 1623 int i; 1624 1625 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 1626 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 1627 return; 1628 1629 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0); 1630 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 1631 1632 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc)); 1633 1634 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1635 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 1636 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 1637 vi->ctrl_promisc ? "en" : "dis"); 1638 1639 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti)); 1640 1641 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1642 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 1643 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 1644 vi->ctrl_allmulti ? "en" : "dis"); 1645 1646 uc_count = netdev_uc_count(dev); 1647 mc_count = netdev_mc_count(dev); 1648 /* MAC filter - use one buffer for both lists */ 1649 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 1650 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 1651 mac_data = buf; 1652 if (!buf) 1653 return; 1654 1655 sg_init_table(sg, 2); 1656 1657 /* Store the unicast list and count in the front of the buffer */ 1658 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 1659 i = 0; 1660 netdev_for_each_uc_addr(ha, dev) 1661 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1662 1663 sg_set_buf(&sg[0], mac_data, 1664 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 1665 1666 /* multicast list and count fill the end */ 1667 mac_data = (void *)&mac_data->macs[uc_count][0]; 1668 1669 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 1670 i = 0; 1671 netdev_for_each_mc_addr(ha, dev) 1672 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1673 1674 sg_set_buf(&sg[1], mac_data, 1675 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 1676 1677 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1678 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 1679 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 1680 1681 kfree(buf); 1682 } 1683 1684 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 1685 __be16 proto, u16 vid) 1686 { 1687 struct virtnet_info *vi = netdev_priv(dev); 1688 struct scatterlist sg; 1689 1690 vi->ctrl_vid = vid; 1691 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid)); 1692 1693 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1694 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 1695 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 1696 return 0; 1697 } 1698 1699 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 1700 __be16 proto, u16 vid) 1701 { 1702 struct virtnet_info *vi = netdev_priv(dev); 1703 struct scatterlist sg; 1704 1705 vi->ctrl_vid = vid; 1706 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid)); 1707 1708 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1709 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 1710 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 1711 return 0; 1712 } 1713 1714 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu) 1715 { 1716 int i; 1717 1718 if (vi->affinity_hint_set) { 1719 for (i = 0; i < vi->max_queue_pairs; i++) { 1720 virtqueue_set_affinity(vi->rq[i].vq, -1); 1721 virtqueue_set_affinity(vi->sq[i].vq, -1); 1722 } 1723 1724 vi->affinity_hint_set = false; 1725 } 1726 } 1727 1728 static void virtnet_set_affinity(struct virtnet_info *vi) 1729 { 1730 int i; 1731 int cpu; 1732 1733 /* In multiqueue mode, when the number of cpu is equal to the number of 1734 * queue pairs, we let the queue pairs to be private to one cpu by 1735 * setting the affinity hint to eliminate the contention. 1736 */ 1737 if (vi->curr_queue_pairs == 1 || 1738 vi->max_queue_pairs != num_online_cpus()) { 1739 virtnet_clean_affinity(vi, -1); 1740 return; 1741 } 1742 1743 i = 0; 1744 for_each_online_cpu(cpu) { 1745 virtqueue_set_affinity(vi->rq[i].vq, cpu); 1746 virtqueue_set_affinity(vi->sq[i].vq, cpu); 1747 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i); 1748 i++; 1749 } 1750 1751 vi->affinity_hint_set = true; 1752 } 1753 1754 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 1755 { 1756 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1757 node); 1758 virtnet_set_affinity(vi); 1759 return 0; 1760 } 1761 1762 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 1763 { 1764 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1765 node_dead); 1766 virtnet_set_affinity(vi); 1767 return 0; 1768 } 1769 1770 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 1771 { 1772 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1773 node); 1774 1775 virtnet_clean_affinity(vi, cpu); 1776 return 0; 1777 } 1778 1779 static enum cpuhp_state virtionet_online; 1780 1781 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 1782 { 1783 int ret; 1784 1785 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 1786 if (ret) 1787 return ret; 1788 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1789 &vi->node_dead); 1790 if (!ret) 1791 return ret; 1792 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1793 return ret; 1794 } 1795 1796 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 1797 { 1798 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1799 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1800 &vi->node_dead); 1801 } 1802 1803 static void virtnet_get_ringparam(struct net_device *dev, 1804 struct ethtool_ringparam *ring) 1805 { 1806 struct virtnet_info *vi = netdev_priv(dev); 1807 1808 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq); 1809 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq); 1810 ring->rx_pending = ring->rx_max_pending; 1811 ring->tx_pending = ring->tx_max_pending; 1812 } 1813 1814 1815 static void virtnet_get_drvinfo(struct net_device *dev, 1816 struct ethtool_drvinfo *info) 1817 { 1818 struct virtnet_info *vi = netdev_priv(dev); 1819 struct virtio_device *vdev = vi->vdev; 1820 1821 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 1822 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 1823 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 1824 1825 } 1826 1827 /* TODO: Eliminate OOO packets during switching */ 1828 static int virtnet_set_channels(struct net_device *dev, 1829 struct ethtool_channels *channels) 1830 { 1831 struct virtnet_info *vi = netdev_priv(dev); 1832 u16 queue_pairs = channels->combined_count; 1833 int err; 1834 1835 /* We don't support separate rx/tx channels. 1836 * We don't allow setting 'other' channels. 1837 */ 1838 if (channels->rx_count || channels->tx_count || channels->other_count) 1839 return -EINVAL; 1840 1841 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 1842 return -EINVAL; 1843 1844 /* For now we don't support modifying channels while XDP is loaded 1845 * also when XDP is loaded all RX queues have XDP programs so we only 1846 * need to check a single RX queue. 1847 */ 1848 if (vi->rq[0].xdp_prog) 1849 return -EINVAL; 1850 1851 get_online_cpus(); 1852 err = _virtnet_set_queues(vi, queue_pairs); 1853 if (!err) { 1854 netif_set_real_num_tx_queues(dev, queue_pairs); 1855 netif_set_real_num_rx_queues(dev, queue_pairs); 1856 1857 virtnet_set_affinity(vi); 1858 } 1859 put_online_cpus(); 1860 1861 return err; 1862 } 1863 1864 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 1865 { 1866 struct virtnet_info *vi = netdev_priv(dev); 1867 char *p = (char *)data; 1868 unsigned int i, j; 1869 1870 switch (stringset) { 1871 case ETH_SS_STATS: 1872 for (i = 0; i < vi->curr_queue_pairs; i++) { 1873 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 1874 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s", 1875 i, virtnet_rq_stats_desc[j].desc); 1876 p += ETH_GSTRING_LEN; 1877 } 1878 } 1879 1880 for (i = 0; i < vi->curr_queue_pairs; i++) { 1881 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 1882 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s", 1883 i, virtnet_sq_stats_desc[j].desc); 1884 p += ETH_GSTRING_LEN; 1885 } 1886 } 1887 break; 1888 } 1889 } 1890 1891 static int virtnet_get_sset_count(struct net_device *dev, int sset) 1892 { 1893 struct virtnet_info *vi = netdev_priv(dev); 1894 1895 switch (sset) { 1896 case ETH_SS_STATS: 1897 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 1898 VIRTNET_SQ_STATS_LEN); 1899 default: 1900 return -EOPNOTSUPP; 1901 } 1902 } 1903 1904 static void virtnet_get_ethtool_stats(struct net_device *dev, 1905 struct ethtool_stats *stats, u64 *data) 1906 { 1907 struct virtnet_info *vi = netdev_priv(dev); 1908 unsigned int idx = 0, start, i, j; 1909 const u8 *stats_base; 1910 size_t offset; 1911 1912 for (i = 0; i < vi->curr_queue_pairs; i++) { 1913 struct receive_queue *rq = &vi->rq[i]; 1914 1915 stats_base = (u8 *)&rq->stats; 1916 do { 1917 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1918 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 1919 offset = virtnet_rq_stats_desc[j].offset; 1920 data[idx + j] = *(u64 *)(stats_base + offset); 1921 } 1922 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 1923 idx += VIRTNET_RQ_STATS_LEN; 1924 } 1925 1926 for (i = 0; i < vi->curr_queue_pairs; i++) { 1927 struct send_queue *sq = &vi->sq[i]; 1928 1929 stats_base = (u8 *)&sq->stats; 1930 do { 1931 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1932 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 1933 offset = virtnet_sq_stats_desc[j].offset; 1934 data[idx + j] = *(u64 *)(stats_base + offset); 1935 } 1936 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1937 idx += VIRTNET_SQ_STATS_LEN; 1938 } 1939 } 1940 1941 static void virtnet_get_channels(struct net_device *dev, 1942 struct ethtool_channels *channels) 1943 { 1944 struct virtnet_info *vi = netdev_priv(dev); 1945 1946 channels->combined_count = vi->curr_queue_pairs; 1947 channels->max_combined = vi->max_queue_pairs; 1948 channels->max_other = 0; 1949 channels->rx_count = 0; 1950 channels->tx_count = 0; 1951 channels->other_count = 0; 1952 } 1953 1954 /* Check if the user is trying to change anything besides speed/duplex */ 1955 static bool 1956 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd) 1957 { 1958 struct ethtool_link_ksettings diff1 = *cmd; 1959 struct ethtool_link_ksettings diff2 = {}; 1960 1961 /* cmd is always set so we need to clear it, validate the port type 1962 * and also without autonegotiation we can ignore advertising 1963 */ 1964 diff1.base.speed = 0; 1965 diff2.base.port = PORT_OTHER; 1966 ethtool_link_ksettings_zero_link_mode(&diff1, advertising); 1967 diff1.base.duplex = 0; 1968 diff1.base.cmd = 0; 1969 diff1.base.link_mode_masks_nwords = 0; 1970 1971 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) && 1972 bitmap_empty(diff1.link_modes.supported, 1973 __ETHTOOL_LINK_MODE_MASK_NBITS) && 1974 bitmap_empty(diff1.link_modes.advertising, 1975 __ETHTOOL_LINK_MODE_MASK_NBITS) && 1976 bitmap_empty(diff1.link_modes.lp_advertising, 1977 __ETHTOOL_LINK_MODE_MASK_NBITS); 1978 } 1979 1980 static int virtnet_set_link_ksettings(struct net_device *dev, 1981 const struct ethtool_link_ksettings *cmd) 1982 { 1983 struct virtnet_info *vi = netdev_priv(dev); 1984 u32 speed; 1985 1986 speed = cmd->base.speed; 1987 /* don't allow custom speed and duplex */ 1988 if (!ethtool_validate_speed(speed) || 1989 !ethtool_validate_duplex(cmd->base.duplex) || 1990 !virtnet_validate_ethtool_cmd(cmd)) 1991 return -EINVAL; 1992 vi->speed = speed; 1993 vi->duplex = cmd->base.duplex; 1994 1995 return 0; 1996 } 1997 1998 static int virtnet_get_link_ksettings(struct net_device *dev, 1999 struct ethtool_link_ksettings *cmd) 2000 { 2001 struct virtnet_info *vi = netdev_priv(dev); 2002 2003 cmd->base.speed = vi->speed; 2004 cmd->base.duplex = vi->duplex; 2005 cmd->base.port = PORT_OTHER; 2006 2007 return 0; 2008 } 2009 2010 static void virtnet_init_settings(struct net_device *dev) 2011 { 2012 struct virtnet_info *vi = netdev_priv(dev); 2013 2014 vi->speed = SPEED_UNKNOWN; 2015 vi->duplex = DUPLEX_UNKNOWN; 2016 } 2017 2018 static void virtnet_update_settings(struct virtnet_info *vi) 2019 { 2020 u32 speed; 2021 u8 duplex; 2022 2023 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 2024 return; 2025 2026 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config, 2027 speed)); 2028 if (ethtool_validate_speed(speed)) 2029 vi->speed = speed; 2030 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config, 2031 duplex)); 2032 if (ethtool_validate_duplex(duplex)) 2033 vi->duplex = duplex; 2034 } 2035 2036 static const struct ethtool_ops virtnet_ethtool_ops = { 2037 .get_drvinfo = virtnet_get_drvinfo, 2038 .get_link = ethtool_op_get_link, 2039 .get_ringparam = virtnet_get_ringparam, 2040 .get_strings = virtnet_get_strings, 2041 .get_sset_count = virtnet_get_sset_count, 2042 .get_ethtool_stats = virtnet_get_ethtool_stats, 2043 .set_channels = virtnet_set_channels, 2044 .get_channels = virtnet_get_channels, 2045 .get_ts_info = ethtool_op_get_ts_info, 2046 .get_link_ksettings = virtnet_get_link_ksettings, 2047 .set_link_ksettings = virtnet_set_link_ksettings, 2048 }; 2049 2050 static void virtnet_freeze_down(struct virtio_device *vdev) 2051 { 2052 struct virtnet_info *vi = vdev->priv; 2053 int i; 2054 2055 /* Make sure no work handler is accessing the device */ 2056 flush_work(&vi->config_work); 2057 2058 netif_device_detach(vi->dev); 2059 netif_tx_disable(vi->dev); 2060 cancel_delayed_work_sync(&vi->refill); 2061 2062 if (netif_running(vi->dev)) { 2063 for (i = 0; i < vi->max_queue_pairs; i++) { 2064 napi_disable(&vi->rq[i].napi); 2065 virtnet_napi_tx_disable(&vi->sq[i].napi); 2066 } 2067 } 2068 } 2069 2070 static int init_vqs(struct virtnet_info *vi); 2071 2072 static int virtnet_restore_up(struct virtio_device *vdev) 2073 { 2074 struct virtnet_info *vi = vdev->priv; 2075 int err, i; 2076 2077 err = init_vqs(vi); 2078 if (err) 2079 return err; 2080 2081 virtio_device_ready(vdev); 2082 2083 if (netif_running(vi->dev)) { 2084 for (i = 0; i < vi->curr_queue_pairs; i++) 2085 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 2086 schedule_delayed_work(&vi->refill, 0); 2087 2088 for (i = 0; i < vi->max_queue_pairs; i++) { 2089 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2090 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 2091 &vi->sq[i].napi); 2092 } 2093 } 2094 2095 netif_device_attach(vi->dev); 2096 return err; 2097 } 2098 2099 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 2100 { 2101 struct scatterlist sg; 2102 vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads); 2103 2104 sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads)); 2105 2106 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 2107 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 2108 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n"); 2109 return -EINVAL; 2110 } 2111 2112 return 0; 2113 } 2114 2115 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 2116 { 2117 u64 offloads = 0; 2118 2119 if (!vi->guest_offloads) 2120 return 0; 2121 2122 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2123 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2124 2125 return virtnet_set_guest_offloads(vi, offloads); 2126 } 2127 2128 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 2129 { 2130 u64 offloads = vi->guest_offloads; 2131 2132 if (!vi->guest_offloads) 2133 return 0; 2134 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2135 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2136 2137 return virtnet_set_guest_offloads(vi, offloads); 2138 } 2139 2140 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 2141 struct netlink_ext_ack *extack) 2142 { 2143 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr); 2144 struct virtnet_info *vi = netdev_priv(dev); 2145 struct bpf_prog *old_prog; 2146 u16 xdp_qp = 0, curr_qp; 2147 int i, err; 2148 2149 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 2150 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 2151 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 2152 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 2153 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) { 2154 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first"); 2155 return -EOPNOTSUPP; 2156 } 2157 2158 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 2159 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 2160 return -EINVAL; 2161 } 2162 2163 if (dev->mtu > max_sz) { 2164 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP"); 2165 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz); 2166 return -EINVAL; 2167 } 2168 2169 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 2170 if (prog) 2171 xdp_qp = nr_cpu_ids; 2172 2173 /* XDP requires extra queues for XDP_TX */ 2174 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 2175 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available"); 2176 netdev_warn(dev, "request %i queues but max is %i\n", 2177 curr_qp + xdp_qp, vi->max_queue_pairs); 2178 return -ENOMEM; 2179 } 2180 2181 if (prog) { 2182 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1); 2183 if (IS_ERR(prog)) 2184 return PTR_ERR(prog); 2185 } 2186 2187 /* Make sure NAPI is not using any XDP TX queues for RX. */ 2188 if (netif_running(dev)) 2189 for (i = 0; i < vi->max_queue_pairs; i++) 2190 napi_disable(&vi->rq[i].napi); 2191 2192 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 2193 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 2194 if (err) 2195 goto err; 2196 vi->xdp_queue_pairs = xdp_qp; 2197 2198 for (i = 0; i < vi->max_queue_pairs; i++) { 2199 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2200 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 2201 if (i == 0) { 2202 if (!old_prog) 2203 virtnet_clear_guest_offloads(vi); 2204 if (!prog) 2205 virtnet_restore_guest_offloads(vi); 2206 } 2207 if (old_prog) 2208 bpf_prog_put(old_prog); 2209 if (netif_running(dev)) 2210 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2211 } 2212 2213 return 0; 2214 2215 err: 2216 for (i = 0; i < vi->max_queue_pairs; i++) 2217 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2218 if (prog) 2219 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 2220 return err; 2221 } 2222 2223 static u32 virtnet_xdp_query(struct net_device *dev) 2224 { 2225 struct virtnet_info *vi = netdev_priv(dev); 2226 const struct bpf_prog *xdp_prog; 2227 int i; 2228 2229 for (i = 0; i < vi->max_queue_pairs; i++) { 2230 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2231 if (xdp_prog) 2232 return xdp_prog->aux->id; 2233 } 2234 return 0; 2235 } 2236 2237 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 2238 { 2239 switch (xdp->command) { 2240 case XDP_SETUP_PROG: 2241 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 2242 case XDP_QUERY_PROG: 2243 xdp->prog_id = virtnet_xdp_query(dev); 2244 xdp->prog_attached = !!xdp->prog_id; 2245 return 0; 2246 default: 2247 return -EINVAL; 2248 } 2249 } 2250 2251 static const struct net_device_ops virtnet_netdev = { 2252 .ndo_open = virtnet_open, 2253 .ndo_stop = virtnet_close, 2254 .ndo_start_xmit = start_xmit, 2255 .ndo_validate_addr = eth_validate_addr, 2256 .ndo_set_mac_address = virtnet_set_mac_address, 2257 .ndo_set_rx_mode = virtnet_set_rx_mode, 2258 .ndo_get_stats64 = virtnet_stats, 2259 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 2260 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 2261 #ifdef CONFIG_NET_POLL_CONTROLLER 2262 .ndo_poll_controller = virtnet_netpoll, 2263 #endif 2264 .ndo_bpf = virtnet_xdp, 2265 .ndo_xdp_xmit = virtnet_xdp_xmit, 2266 .ndo_xdp_flush = virtnet_xdp_flush, 2267 .ndo_features_check = passthru_features_check, 2268 }; 2269 2270 static void virtnet_config_changed_work(struct work_struct *work) 2271 { 2272 struct virtnet_info *vi = 2273 container_of(work, struct virtnet_info, config_work); 2274 u16 v; 2275 2276 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 2277 struct virtio_net_config, status, &v) < 0) 2278 return; 2279 2280 if (v & VIRTIO_NET_S_ANNOUNCE) { 2281 netdev_notify_peers(vi->dev); 2282 virtnet_ack_link_announce(vi); 2283 } 2284 2285 /* Ignore unknown (future) status bits */ 2286 v &= VIRTIO_NET_S_LINK_UP; 2287 2288 if (vi->status == v) 2289 return; 2290 2291 vi->status = v; 2292 2293 if (vi->status & VIRTIO_NET_S_LINK_UP) { 2294 virtnet_update_settings(vi); 2295 netif_carrier_on(vi->dev); 2296 netif_tx_wake_all_queues(vi->dev); 2297 } else { 2298 netif_carrier_off(vi->dev); 2299 netif_tx_stop_all_queues(vi->dev); 2300 } 2301 } 2302 2303 static void virtnet_config_changed(struct virtio_device *vdev) 2304 { 2305 struct virtnet_info *vi = vdev->priv; 2306 2307 schedule_work(&vi->config_work); 2308 } 2309 2310 static void virtnet_free_queues(struct virtnet_info *vi) 2311 { 2312 int i; 2313 2314 for (i = 0; i < vi->max_queue_pairs; i++) { 2315 napi_hash_del(&vi->rq[i].napi); 2316 netif_napi_del(&vi->rq[i].napi); 2317 netif_napi_del(&vi->sq[i].napi); 2318 } 2319 2320 /* We called napi_hash_del() before netif_napi_del(), 2321 * we need to respect an RCU grace period before freeing vi->rq 2322 */ 2323 synchronize_net(); 2324 2325 kfree(vi->rq); 2326 kfree(vi->sq); 2327 } 2328 2329 static void _free_receive_bufs(struct virtnet_info *vi) 2330 { 2331 struct bpf_prog *old_prog; 2332 int i; 2333 2334 for (i = 0; i < vi->max_queue_pairs; i++) { 2335 while (vi->rq[i].pages) 2336 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 2337 2338 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2339 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 2340 if (old_prog) 2341 bpf_prog_put(old_prog); 2342 } 2343 } 2344 2345 static void free_receive_bufs(struct virtnet_info *vi) 2346 { 2347 rtnl_lock(); 2348 _free_receive_bufs(vi); 2349 rtnl_unlock(); 2350 } 2351 2352 static void free_receive_page_frags(struct virtnet_info *vi) 2353 { 2354 int i; 2355 for (i = 0; i < vi->max_queue_pairs; i++) 2356 if (vi->rq[i].alloc_frag.page) 2357 put_page(vi->rq[i].alloc_frag.page); 2358 } 2359 2360 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) 2361 { 2362 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs)) 2363 return false; 2364 else if (q < vi->curr_queue_pairs) 2365 return true; 2366 else 2367 return false; 2368 } 2369 2370 static void free_unused_bufs(struct virtnet_info *vi) 2371 { 2372 void *buf; 2373 int i; 2374 2375 for (i = 0; i < vi->max_queue_pairs; i++) { 2376 struct virtqueue *vq = vi->sq[i].vq; 2377 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2378 if (!is_xdp_raw_buffer_queue(vi, i)) 2379 dev_kfree_skb(buf); 2380 else 2381 put_page(virt_to_head_page(buf)); 2382 } 2383 } 2384 2385 for (i = 0; i < vi->max_queue_pairs; i++) { 2386 struct virtqueue *vq = vi->rq[i].vq; 2387 2388 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2389 if (vi->mergeable_rx_bufs) { 2390 put_page(virt_to_head_page(buf)); 2391 } else if (vi->big_packets) { 2392 give_pages(&vi->rq[i], buf); 2393 } else { 2394 put_page(virt_to_head_page(buf)); 2395 } 2396 } 2397 } 2398 } 2399 2400 static void virtnet_del_vqs(struct virtnet_info *vi) 2401 { 2402 struct virtio_device *vdev = vi->vdev; 2403 2404 virtnet_clean_affinity(vi, -1); 2405 2406 vdev->config->del_vqs(vdev); 2407 2408 virtnet_free_queues(vi); 2409 } 2410 2411 /* How large should a single buffer be so a queue full of these can fit at 2412 * least one full packet? 2413 * Logic below assumes the mergeable buffer header is used. 2414 */ 2415 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 2416 { 2417 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2418 unsigned int rq_size = virtqueue_get_vring_size(vq); 2419 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 2420 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 2421 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 2422 2423 return max(max(min_buf_len, hdr_len) - hdr_len, 2424 (unsigned int)GOOD_PACKET_LEN); 2425 } 2426 2427 static int virtnet_find_vqs(struct virtnet_info *vi) 2428 { 2429 vq_callback_t **callbacks; 2430 struct virtqueue **vqs; 2431 int ret = -ENOMEM; 2432 int i, total_vqs; 2433 const char **names; 2434 bool *ctx; 2435 2436 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 2437 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 2438 * possible control vq. 2439 */ 2440 total_vqs = vi->max_queue_pairs * 2 + 2441 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 2442 2443 /* Allocate space for find_vqs parameters */ 2444 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL); 2445 if (!vqs) 2446 goto err_vq; 2447 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL); 2448 if (!callbacks) 2449 goto err_callback; 2450 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL); 2451 if (!names) 2452 goto err_names; 2453 if (!vi->big_packets || vi->mergeable_rx_bufs) { 2454 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL); 2455 if (!ctx) 2456 goto err_ctx; 2457 } else { 2458 ctx = NULL; 2459 } 2460 2461 /* Parameters for control virtqueue, if any */ 2462 if (vi->has_cvq) { 2463 callbacks[total_vqs - 1] = NULL; 2464 names[total_vqs - 1] = "control"; 2465 } 2466 2467 /* Allocate/initialize parameters for send/receive virtqueues */ 2468 for (i = 0; i < vi->max_queue_pairs; i++) { 2469 callbacks[rxq2vq(i)] = skb_recv_done; 2470 callbacks[txq2vq(i)] = skb_xmit_done; 2471 sprintf(vi->rq[i].name, "input.%d", i); 2472 sprintf(vi->sq[i].name, "output.%d", i); 2473 names[rxq2vq(i)] = vi->rq[i].name; 2474 names[txq2vq(i)] = vi->sq[i].name; 2475 if (ctx) 2476 ctx[rxq2vq(i)] = true; 2477 } 2478 2479 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks, 2480 names, ctx, NULL); 2481 if (ret) 2482 goto err_find; 2483 2484 if (vi->has_cvq) { 2485 vi->cvq = vqs[total_vqs - 1]; 2486 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 2487 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 2488 } 2489 2490 for (i = 0; i < vi->max_queue_pairs; i++) { 2491 vi->rq[i].vq = vqs[rxq2vq(i)]; 2492 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 2493 vi->sq[i].vq = vqs[txq2vq(i)]; 2494 } 2495 2496 kfree(names); 2497 kfree(callbacks); 2498 kfree(vqs); 2499 kfree(ctx); 2500 2501 return 0; 2502 2503 err_find: 2504 kfree(ctx); 2505 err_ctx: 2506 kfree(names); 2507 err_names: 2508 kfree(callbacks); 2509 err_callback: 2510 kfree(vqs); 2511 err_vq: 2512 return ret; 2513 } 2514 2515 static int virtnet_alloc_queues(struct virtnet_info *vi) 2516 { 2517 int i; 2518 2519 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL); 2520 if (!vi->sq) 2521 goto err_sq; 2522 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL); 2523 if (!vi->rq) 2524 goto err_rq; 2525 2526 INIT_DELAYED_WORK(&vi->refill, refill_work); 2527 for (i = 0; i < vi->max_queue_pairs; i++) { 2528 vi->rq[i].pages = NULL; 2529 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll, 2530 napi_weight); 2531 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx, 2532 napi_tx ? napi_weight : 0); 2533 2534 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 2535 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 2536 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 2537 2538 u64_stats_init(&vi->rq[i].stats.syncp); 2539 u64_stats_init(&vi->sq[i].stats.syncp); 2540 } 2541 2542 return 0; 2543 2544 err_rq: 2545 kfree(vi->sq); 2546 err_sq: 2547 return -ENOMEM; 2548 } 2549 2550 static int init_vqs(struct virtnet_info *vi) 2551 { 2552 int ret; 2553 2554 /* Allocate send & receive queues */ 2555 ret = virtnet_alloc_queues(vi); 2556 if (ret) 2557 goto err; 2558 2559 ret = virtnet_find_vqs(vi); 2560 if (ret) 2561 goto err_free; 2562 2563 get_online_cpus(); 2564 virtnet_set_affinity(vi); 2565 put_online_cpus(); 2566 2567 return 0; 2568 2569 err_free: 2570 virtnet_free_queues(vi); 2571 err: 2572 return ret; 2573 } 2574 2575 #ifdef CONFIG_SYSFS 2576 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 2577 char *buf) 2578 { 2579 struct virtnet_info *vi = netdev_priv(queue->dev); 2580 unsigned int queue_index = get_netdev_rx_queue_index(queue); 2581 struct ewma_pkt_len *avg; 2582 2583 BUG_ON(queue_index >= vi->max_queue_pairs); 2584 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 2585 return sprintf(buf, "%u\n", 2586 get_mergeable_buf_len(&vi->rq[queue_index], avg)); 2587 } 2588 2589 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 2590 __ATTR_RO(mergeable_rx_buffer_size); 2591 2592 static struct attribute *virtio_net_mrg_rx_attrs[] = { 2593 &mergeable_rx_buffer_size_attribute.attr, 2594 NULL 2595 }; 2596 2597 static const struct attribute_group virtio_net_mrg_rx_group = { 2598 .name = "virtio_net", 2599 .attrs = virtio_net_mrg_rx_attrs 2600 }; 2601 #endif 2602 2603 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 2604 unsigned int fbit, 2605 const char *fname, const char *dname) 2606 { 2607 if (!virtio_has_feature(vdev, fbit)) 2608 return false; 2609 2610 dev_err(&vdev->dev, "device advertises feature %s but not %s", 2611 fname, dname); 2612 2613 return true; 2614 } 2615 2616 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 2617 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 2618 2619 static bool virtnet_validate_features(struct virtio_device *vdev) 2620 { 2621 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 2622 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 2623 "VIRTIO_NET_F_CTRL_VQ") || 2624 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 2625 "VIRTIO_NET_F_CTRL_VQ") || 2626 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 2627 "VIRTIO_NET_F_CTRL_VQ") || 2628 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 2629 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 2630 "VIRTIO_NET_F_CTRL_VQ"))) { 2631 return false; 2632 } 2633 2634 return true; 2635 } 2636 2637 #define MIN_MTU ETH_MIN_MTU 2638 #define MAX_MTU ETH_MAX_MTU 2639 2640 static int virtnet_validate(struct virtio_device *vdev) 2641 { 2642 if (!vdev->config->get) { 2643 dev_err(&vdev->dev, "%s failure: config access disabled\n", 2644 __func__); 2645 return -EINVAL; 2646 } 2647 2648 if (!virtnet_validate_features(vdev)) 2649 return -EINVAL; 2650 2651 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2652 int mtu = virtio_cread16(vdev, 2653 offsetof(struct virtio_net_config, 2654 mtu)); 2655 if (mtu < MIN_MTU) 2656 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 2657 } 2658 2659 return 0; 2660 } 2661 2662 static int virtnet_probe(struct virtio_device *vdev) 2663 { 2664 int i, err = -ENOMEM; 2665 struct net_device *dev; 2666 struct virtnet_info *vi; 2667 u16 max_queue_pairs; 2668 int mtu; 2669 2670 /* Find if host supports multiqueue virtio_net device */ 2671 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ, 2672 struct virtio_net_config, 2673 max_virtqueue_pairs, &max_queue_pairs); 2674 2675 /* We need at least 2 queue's */ 2676 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 2677 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 2678 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2679 max_queue_pairs = 1; 2680 2681 /* Allocate ourselves a network device with room for our info */ 2682 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 2683 if (!dev) 2684 return -ENOMEM; 2685 2686 /* Set up network device as normal. */ 2687 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE; 2688 dev->netdev_ops = &virtnet_netdev; 2689 dev->features = NETIF_F_HIGHDMA; 2690 2691 dev->ethtool_ops = &virtnet_ethtool_ops; 2692 SET_NETDEV_DEV(dev, &vdev->dev); 2693 2694 /* Do we support "hardware" checksums? */ 2695 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 2696 /* This opens up the world of extra features. */ 2697 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2698 if (csum) 2699 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2700 2701 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 2702 dev->hw_features |= NETIF_F_TSO 2703 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 2704 } 2705 /* Individual feature bits: what can host handle? */ 2706 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 2707 dev->hw_features |= NETIF_F_TSO; 2708 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 2709 dev->hw_features |= NETIF_F_TSO6; 2710 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 2711 dev->hw_features |= NETIF_F_TSO_ECN; 2712 2713 dev->features |= NETIF_F_GSO_ROBUST; 2714 2715 if (gso) 2716 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 2717 /* (!csum && gso) case will be fixed by register_netdev() */ 2718 } 2719 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 2720 dev->features |= NETIF_F_RXCSUM; 2721 2722 dev->vlan_features = dev->features; 2723 2724 /* MTU range: 68 - 65535 */ 2725 dev->min_mtu = MIN_MTU; 2726 dev->max_mtu = MAX_MTU; 2727 2728 /* Configuration may specify what MAC to use. Otherwise random. */ 2729 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) 2730 virtio_cread_bytes(vdev, 2731 offsetof(struct virtio_net_config, mac), 2732 dev->dev_addr, dev->addr_len); 2733 else 2734 eth_hw_addr_random(dev); 2735 2736 /* Set up our device-specific information */ 2737 vi = netdev_priv(dev); 2738 vi->dev = dev; 2739 vi->vdev = vdev; 2740 vdev->priv = vi; 2741 2742 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 2743 2744 /* If we can receive ANY GSO packets, we must allocate large ones. */ 2745 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 2746 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) || 2747 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) || 2748 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO)) 2749 vi->big_packets = true; 2750 2751 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) 2752 vi->mergeable_rx_bufs = true; 2753 2754 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 2755 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2756 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2757 else 2758 vi->hdr_len = sizeof(struct virtio_net_hdr); 2759 2760 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 2761 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2762 vi->any_header_sg = true; 2763 2764 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2765 vi->has_cvq = true; 2766 2767 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2768 mtu = virtio_cread16(vdev, 2769 offsetof(struct virtio_net_config, 2770 mtu)); 2771 if (mtu < dev->min_mtu) { 2772 /* Should never trigger: MTU was previously validated 2773 * in virtnet_validate. 2774 */ 2775 dev_err(&vdev->dev, "device MTU appears to have changed " 2776 "it is now %d < %d", mtu, dev->min_mtu); 2777 goto free; 2778 } 2779 2780 dev->mtu = mtu; 2781 dev->max_mtu = mtu; 2782 2783 /* TODO: size buffers correctly in this case. */ 2784 if (dev->mtu > ETH_DATA_LEN) 2785 vi->big_packets = true; 2786 } 2787 2788 if (vi->any_header_sg) 2789 dev->needed_headroom = vi->hdr_len; 2790 2791 /* Enable multiqueue by default */ 2792 if (num_online_cpus() >= max_queue_pairs) 2793 vi->curr_queue_pairs = max_queue_pairs; 2794 else 2795 vi->curr_queue_pairs = num_online_cpus(); 2796 vi->max_queue_pairs = max_queue_pairs; 2797 2798 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 2799 err = init_vqs(vi); 2800 if (err) 2801 goto free; 2802 2803 #ifdef CONFIG_SYSFS 2804 if (vi->mergeable_rx_bufs) 2805 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 2806 #endif 2807 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 2808 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 2809 2810 virtnet_init_settings(dev); 2811 2812 err = register_netdev(dev); 2813 if (err) { 2814 pr_debug("virtio_net: registering device failed\n"); 2815 goto free_vqs; 2816 } 2817 2818 virtio_device_ready(vdev); 2819 2820 err = virtnet_cpu_notif_add(vi); 2821 if (err) { 2822 pr_debug("virtio_net: registering cpu notifier failed\n"); 2823 goto free_unregister_netdev; 2824 } 2825 2826 virtnet_set_queues(vi, vi->curr_queue_pairs); 2827 2828 /* Assume link up if device can't report link status, 2829 otherwise get link status from config. */ 2830 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 2831 netif_carrier_off(dev); 2832 schedule_work(&vi->config_work); 2833 } else { 2834 vi->status = VIRTIO_NET_S_LINK_UP; 2835 virtnet_update_settings(vi); 2836 netif_carrier_on(dev); 2837 } 2838 2839 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 2840 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 2841 set_bit(guest_offloads[i], &vi->guest_offloads); 2842 2843 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 2844 dev->name, max_queue_pairs); 2845 2846 return 0; 2847 2848 free_unregister_netdev: 2849 vi->vdev->config->reset(vdev); 2850 2851 unregister_netdev(dev); 2852 free_vqs: 2853 cancel_delayed_work_sync(&vi->refill); 2854 free_receive_page_frags(vi); 2855 virtnet_del_vqs(vi); 2856 free: 2857 free_netdev(dev); 2858 return err; 2859 } 2860 2861 static void remove_vq_common(struct virtnet_info *vi) 2862 { 2863 vi->vdev->config->reset(vi->vdev); 2864 2865 /* Free unused buffers in both send and recv, if any. */ 2866 free_unused_bufs(vi); 2867 2868 free_receive_bufs(vi); 2869 2870 free_receive_page_frags(vi); 2871 2872 virtnet_del_vqs(vi); 2873 } 2874 2875 static void virtnet_remove(struct virtio_device *vdev) 2876 { 2877 struct virtnet_info *vi = vdev->priv; 2878 2879 virtnet_cpu_notif_remove(vi); 2880 2881 /* Make sure no work handler is accessing the device. */ 2882 flush_work(&vi->config_work); 2883 2884 unregister_netdev(vi->dev); 2885 2886 remove_vq_common(vi); 2887 2888 free_netdev(vi->dev); 2889 } 2890 2891 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 2892 { 2893 struct virtnet_info *vi = vdev->priv; 2894 2895 virtnet_cpu_notif_remove(vi); 2896 virtnet_freeze_down(vdev); 2897 remove_vq_common(vi); 2898 2899 return 0; 2900 } 2901 2902 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 2903 { 2904 struct virtnet_info *vi = vdev->priv; 2905 int err; 2906 2907 err = virtnet_restore_up(vdev); 2908 if (err) 2909 return err; 2910 virtnet_set_queues(vi, vi->curr_queue_pairs); 2911 2912 err = virtnet_cpu_notif_add(vi); 2913 if (err) 2914 return err; 2915 2916 return 0; 2917 } 2918 2919 static struct virtio_device_id id_table[] = { 2920 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 2921 { 0 }, 2922 }; 2923 2924 #define VIRTNET_FEATURES \ 2925 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 2926 VIRTIO_NET_F_MAC, \ 2927 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 2928 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 2929 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 2930 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 2931 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 2932 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 2933 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 2934 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 2935 VIRTIO_NET_F_SPEED_DUPLEX 2936 2937 static unsigned int features[] = { 2938 VIRTNET_FEATURES, 2939 }; 2940 2941 static unsigned int features_legacy[] = { 2942 VIRTNET_FEATURES, 2943 VIRTIO_NET_F_GSO, 2944 VIRTIO_F_ANY_LAYOUT, 2945 }; 2946 2947 static struct virtio_driver virtio_net_driver = { 2948 .feature_table = features, 2949 .feature_table_size = ARRAY_SIZE(features), 2950 .feature_table_legacy = features_legacy, 2951 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 2952 .driver.name = KBUILD_MODNAME, 2953 .driver.owner = THIS_MODULE, 2954 .id_table = id_table, 2955 .validate = virtnet_validate, 2956 .probe = virtnet_probe, 2957 .remove = virtnet_remove, 2958 .config_changed = virtnet_config_changed, 2959 #ifdef CONFIG_PM_SLEEP 2960 .freeze = virtnet_freeze, 2961 .restore = virtnet_restore, 2962 #endif 2963 }; 2964 2965 static __init int virtio_net_driver_init(void) 2966 { 2967 int ret; 2968 2969 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 2970 virtnet_cpu_online, 2971 virtnet_cpu_down_prep); 2972 if (ret < 0) 2973 goto out; 2974 virtionet_online = ret; 2975 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 2976 NULL, virtnet_cpu_dead); 2977 if (ret) 2978 goto err_dead; 2979 2980 ret = register_virtio_driver(&virtio_net_driver); 2981 if (ret) 2982 goto err_virtio; 2983 return 0; 2984 err_virtio: 2985 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 2986 err_dead: 2987 cpuhp_remove_multi_state(virtionet_online); 2988 out: 2989 return ret; 2990 } 2991 module_init(virtio_net_driver_init); 2992 2993 static __exit void virtio_net_driver_exit(void) 2994 { 2995 unregister_virtio_driver(&virtio_net_driver); 2996 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 2997 cpuhp_remove_multi_state(virtionet_online); 2998 } 2999 module_exit(virtio_net_driver_exit); 3000 3001 MODULE_DEVICE_TABLE(virtio, id_table); 3002 MODULE_DESCRIPTION("Virtio network driver"); 3003 MODULE_LICENSE("GPL"); 3004