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 /* Control VQ buffers: protected by the rtnl lock */ 151 struct control_buf { 152 struct virtio_net_ctrl_hdr hdr; 153 virtio_net_ctrl_ack status; 154 struct virtio_net_ctrl_mq mq; 155 u8 promisc; 156 u8 allmulti; 157 __virtio16 vid; 158 __virtio64 offloads; 159 }; 160 161 struct virtnet_info { 162 struct virtio_device *vdev; 163 struct virtqueue *cvq; 164 struct net_device *dev; 165 struct send_queue *sq; 166 struct receive_queue *rq; 167 unsigned int status; 168 169 /* Max # of queue pairs supported by the device */ 170 u16 max_queue_pairs; 171 172 /* # of queue pairs currently used by the driver */ 173 u16 curr_queue_pairs; 174 175 /* # of XDP queue pairs currently used by the driver */ 176 u16 xdp_queue_pairs; 177 178 /* I like... big packets and I cannot lie! */ 179 bool big_packets; 180 181 /* Host will merge rx buffers for big packets (shake it! shake it!) */ 182 bool mergeable_rx_bufs; 183 184 /* Has control virtqueue */ 185 bool has_cvq; 186 187 /* Host can handle any s/g split between our header and packet data */ 188 bool any_header_sg; 189 190 /* Packet virtio header size */ 191 u8 hdr_len; 192 193 /* Work struct for refilling if we run low on memory. */ 194 struct delayed_work refill; 195 196 /* Work struct for config space updates */ 197 struct work_struct config_work; 198 199 /* Does the affinity hint is set for virtqueues? */ 200 bool affinity_hint_set; 201 202 /* CPU hotplug instances for online & dead */ 203 struct hlist_node node; 204 struct hlist_node node_dead; 205 206 struct control_buf *ctrl; 207 208 /* Ethtool settings */ 209 u8 duplex; 210 u32 speed; 211 212 unsigned long guest_offloads; 213 }; 214 215 struct padded_vnet_hdr { 216 struct virtio_net_hdr_mrg_rxbuf hdr; 217 /* 218 * hdr is in a separate sg buffer, and data sg buffer shares same page 219 * with this header sg. This padding makes next sg 16 byte aligned 220 * after the header. 221 */ 222 char padding[4]; 223 }; 224 225 /* Converting between virtqueue no. and kernel tx/rx queue no. 226 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq 227 */ 228 static int vq2txq(struct virtqueue *vq) 229 { 230 return (vq->index - 1) / 2; 231 } 232 233 static int txq2vq(int txq) 234 { 235 return txq * 2 + 1; 236 } 237 238 static int vq2rxq(struct virtqueue *vq) 239 { 240 return vq->index / 2; 241 } 242 243 static int rxq2vq(int rxq) 244 { 245 return rxq * 2; 246 } 247 248 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb) 249 { 250 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb; 251 } 252 253 /* 254 * private is used to chain pages for big packets, put the whole 255 * most recent used list in the beginning for reuse 256 */ 257 static void give_pages(struct receive_queue *rq, struct page *page) 258 { 259 struct page *end; 260 261 /* Find end of list, sew whole thing into vi->rq.pages. */ 262 for (end = page; end->private; end = (struct page *)end->private); 263 end->private = (unsigned long)rq->pages; 264 rq->pages = page; 265 } 266 267 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask) 268 { 269 struct page *p = rq->pages; 270 271 if (p) { 272 rq->pages = (struct page *)p->private; 273 /* clear private here, it is used to chain pages */ 274 p->private = 0; 275 } else 276 p = alloc_page(gfp_mask); 277 return p; 278 } 279 280 static void virtqueue_napi_schedule(struct napi_struct *napi, 281 struct virtqueue *vq) 282 { 283 if (napi_schedule_prep(napi)) { 284 virtqueue_disable_cb(vq); 285 __napi_schedule(napi); 286 } 287 } 288 289 static void virtqueue_napi_complete(struct napi_struct *napi, 290 struct virtqueue *vq, int processed) 291 { 292 int opaque; 293 294 opaque = virtqueue_enable_cb_prepare(vq); 295 if (napi_complete_done(napi, processed)) { 296 if (unlikely(virtqueue_poll(vq, opaque))) 297 virtqueue_napi_schedule(napi, vq); 298 } else { 299 virtqueue_disable_cb(vq); 300 } 301 } 302 303 static void skb_xmit_done(struct virtqueue *vq) 304 { 305 struct virtnet_info *vi = vq->vdev->priv; 306 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi; 307 308 /* Suppress further interrupts. */ 309 virtqueue_disable_cb(vq); 310 311 if (napi->weight) 312 virtqueue_napi_schedule(napi, vq); 313 else 314 /* We were probably waiting for more output buffers. */ 315 netif_wake_subqueue(vi->dev, vq2txq(vq)); 316 } 317 318 #define MRG_CTX_HEADER_SHIFT 22 319 static void *mergeable_len_to_ctx(unsigned int truesize, 320 unsigned int headroom) 321 { 322 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize); 323 } 324 325 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx) 326 { 327 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT; 328 } 329 330 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx) 331 { 332 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1); 333 } 334 335 /* Called from bottom half context */ 336 static struct sk_buff *page_to_skb(struct virtnet_info *vi, 337 struct receive_queue *rq, 338 struct page *page, unsigned int offset, 339 unsigned int len, unsigned int truesize) 340 { 341 struct sk_buff *skb; 342 struct virtio_net_hdr_mrg_rxbuf *hdr; 343 unsigned int copy, hdr_len, hdr_padded_len; 344 char *p; 345 346 p = page_address(page) + offset; 347 348 /* copy small packet so we can reuse these pages for small data */ 349 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN); 350 if (unlikely(!skb)) 351 return NULL; 352 353 hdr = skb_vnet_hdr(skb); 354 355 hdr_len = vi->hdr_len; 356 if (vi->mergeable_rx_bufs) 357 hdr_padded_len = sizeof(*hdr); 358 else 359 hdr_padded_len = sizeof(struct padded_vnet_hdr); 360 361 memcpy(hdr, p, hdr_len); 362 363 len -= hdr_len; 364 offset += hdr_padded_len; 365 p += hdr_padded_len; 366 367 copy = len; 368 if (copy > skb_tailroom(skb)) 369 copy = skb_tailroom(skb); 370 skb_put_data(skb, p, copy); 371 372 len -= copy; 373 offset += copy; 374 375 if (vi->mergeable_rx_bufs) { 376 if (len) 377 skb_add_rx_frag(skb, 0, page, offset, len, truesize); 378 else 379 put_page(page); 380 return skb; 381 } 382 383 /* 384 * Verify that we can indeed put this data into a skb. 385 * This is here to handle cases when the device erroneously 386 * tries to receive more than is possible. This is usually 387 * the case of a broken device. 388 */ 389 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) { 390 net_dbg_ratelimited("%s: too much data\n", skb->dev->name); 391 dev_kfree_skb(skb); 392 return NULL; 393 } 394 BUG_ON(offset >= PAGE_SIZE); 395 while (len) { 396 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len); 397 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, 398 frag_size, truesize); 399 len -= frag_size; 400 page = (struct page *)page->private; 401 offset = 0; 402 } 403 404 if (page) 405 give_pages(rq, page); 406 407 return skb; 408 } 409 410 static void virtnet_xdp_flush(struct net_device *dev) 411 { 412 struct virtnet_info *vi = netdev_priv(dev); 413 struct send_queue *sq; 414 unsigned int qp; 415 416 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 417 sq = &vi->sq[qp]; 418 419 virtqueue_kick(sq->vq); 420 } 421 422 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 423 struct send_queue *sq, 424 struct xdp_frame *xdpf) 425 { 426 struct virtio_net_hdr_mrg_rxbuf *hdr; 427 int err; 428 429 /* virtqueue want to use data area in-front of packet */ 430 if (unlikely(xdpf->metasize > 0)) 431 return -EOPNOTSUPP; 432 433 if (unlikely(xdpf->headroom < vi->hdr_len)) 434 return -EOVERFLOW; 435 436 /* Make room for virtqueue hdr (also change xdpf->headroom?) */ 437 xdpf->data -= vi->hdr_len; 438 /* Zero header and leave csum up to XDP layers */ 439 hdr = xdpf->data; 440 memset(hdr, 0, vi->hdr_len); 441 xdpf->len += vi->hdr_len; 442 443 sg_init_one(sq->sg, xdpf->data, xdpf->len); 444 445 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC); 446 if (unlikely(err)) 447 return -ENOSPC; /* Caller handle free/refcnt */ 448 449 return 0; 450 } 451 452 static int __virtnet_xdp_tx_xmit(struct virtnet_info *vi, 453 struct xdp_frame *xdpf) 454 { 455 struct xdp_frame *xdpf_sent; 456 struct send_queue *sq; 457 unsigned int len; 458 unsigned int qp; 459 460 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 461 sq = &vi->sq[qp]; 462 463 /* Free up any pending old buffers before queueing new ones. */ 464 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) 465 xdp_return_frame(xdpf_sent); 466 467 return __virtnet_xdp_xmit_one(vi, sq, xdpf); 468 } 469 470 static int virtnet_xdp_xmit(struct net_device *dev, 471 int n, struct xdp_frame **frames) 472 { 473 struct virtnet_info *vi = netdev_priv(dev); 474 struct receive_queue *rq = vi->rq; 475 struct xdp_frame *xdpf_sent; 476 struct bpf_prog *xdp_prog; 477 struct send_queue *sq; 478 unsigned int len; 479 unsigned int qp; 480 int drops = 0; 481 int err; 482 int i; 483 484 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 485 sq = &vi->sq[qp]; 486 487 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this 488 * indicate XDP resources have been successfully allocated. 489 */ 490 xdp_prog = rcu_dereference(rq->xdp_prog); 491 if (!xdp_prog) 492 return -ENXIO; 493 494 /* Free up any pending old buffers before queueing new ones. */ 495 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) 496 xdp_return_frame(xdpf_sent); 497 498 for (i = 0; i < n; i++) { 499 struct xdp_frame *xdpf = frames[i]; 500 501 err = __virtnet_xdp_xmit_one(vi, sq, xdpf); 502 if (err) { 503 xdp_return_frame_rx_napi(xdpf); 504 drops++; 505 } 506 } 507 return n - drops; 508 } 509 510 static unsigned int virtnet_get_headroom(struct virtnet_info *vi) 511 { 512 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0; 513 } 514 515 /* We copy the packet for XDP in the following cases: 516 * 517 * 1) Packet is scattered across multiple rx buffers. 518 * 2) Headroom space is insufficient. 519 * 520 * This is inefficient but it's a temporary condition that 521 * we hit right after XDP is enabled and until queue is refilled 522 * with large buffers with sufficient headroom - so it should affect 523 * at most queue size packets. 524 * Afterwards, the conditions to enable 525 * XDP should preclude the underlying device from sending packets 526 * across multiple buffers (num_buf > 1), and we make sure buffers 527 * have enough headroom. 528 */ 529 static struct page *xdp_linearize_page(struct receive_queue *rq, 530 u16 *num_buf, 531 struct page *p, 532 int offset, 533 int page_off, 534 unsigned int *len) 535 { 536 struct page *page = alloc_page(GFP_ATOMIC); 537 538 if (!page) 539 return NULL; 540 541 memcpy(page_address(page) + page_off, page_address(p) + offset, *len); 542 page_off += *len; 543 544 while (--*num_buf) { 545 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 546 unsigned int buflen; 547 void *buf; 548 int off; 549 550 buf = virtqueue_get_buf(rq->vq, &buflen); 551 if (unlikely(!buf)) 552 goto err_buf; 553 554 p = virt_to_head_page(buf); 555 off = buf - page_address(p); 556 557 /* guard against a misconfigured or uncooperative backend that 558 * is sending packet larger than the MTU. 559 */ 560 if ((page_off + buflen + tailroom) > PAGE_SIZE) { 561 put_page(p); 562 goto err_buf; 563 } 564 565 memcpy(page_address(page) + page_off, 566 page_address(p) + off, buflen); 567 page_off += buflen; 568 put_page(p); 569 } 570 571 /* Headroom does not contribute to packet length */ 572 *len = page_off - VIRTIO_XDP_HEADROOM; 573 return page; 574 err_buf: 575 __free_pages(page, 0); 576 return NULL; 577 } 578 579 static struct sk_buff *receive_small(struct net_device *dev, 580 struct virtnet_info *vi, 581 struct receive_queue *rq, 582 void *buf, void *ctx, 583 unsigned int len, 584 bool *xdp_xmit) 585 { 586 struct sk_buff *skb; 587 struct bpf_prog *xdp_prog; 588 unsigned int xdp_headroom = (unsigned long)ctx; 589 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom; 590 unsigned int headroom = vi->hdr_len + header_offset; 591 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 592 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 593 struct page *page = virt_to_head_page(buf); 594 unsigned int delta = 0; 595 struct page *xdp_page; 596 int err; 597 598 len -= vi->hdr_len; 599 600 rcu_read_lock(); 601 xdp_prog = rcu_dereference(rq->xdp_prog); 602 if (xdp_prog) { 603 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset; 604 struct xdp_frame *xdpf; 605 struct xdp_buff xdp; 606 void *orig_data; 607 u32 act; 608 609 if (unlikely(hdr->hdr.gso_type)) 610 goto err_xdp; 611 612 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) { 613 int offset = buf - page_address(page) + header_offset; 614 unsigned int tlen = len + vi->hdr_len; 615 u16 num_buf = 1; 616 617 xdp_headroom = virtnet_get_headroom(vi); 618 header_offset = VIRTNET_RX_PAD + xdp_headroom; 619 headroom = vi->hdr_len + header_offset; 620 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 621 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 622 xdp_page = xdp_linearize_page(rq, &num_buf, page, 623 offset, header_offset, 624 &tlen); 625 if (!xdp_page) 626 goto err_xdp; 627 628 buf = page_address(xdp_page); 629 put_page(page); 630 page = xdp_page; 631 } 632 633 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len; 634 xdp.data = xdp.data_hard_start + xdp_headroom; 635 xdp_set_data_meta_invalid(&xdp); 636 xdp.data_end = xdp.data + len; 637 xdp.rxq = &rq->xdp_rxq; 638 orig_data = xdp.data; 639 act = bpf_prog_run_xdp(xdp_prog, &xdp); 640 641 switch (act) { 642 case XDP_PASS: 643 /* Recalculate length in case bpf program changed it */ 644 delta = orig_data - xdp.data; 645 len = xdp.data_end - xdp.data; 646 break; 647 case XDP_TX: 648 xdpf = convert_to_xdp_frame(&xdp); 649 if (unlikely(!xdpf)) 650 goto err_xdp; 651 err = __virtnet_xdp_tx_xmit(vi, xdpf); 652 if (unlikely(err)) { 653 trace_xdp_exception(vi->dev, xdp_prog, act); 654 goto err_xdp; 655 } 656 *xdp_xmit = true; 657 rcu_read_unlock(); 658 goto xdp_xmit; 659 case XDP_REDIRECT: 660 err = xdp_do_redirect(dev, &xdp, xdp_prog); 661 if (err) 662 goto err_xdp; 663 *xdp_xmit = true; 664 rcu_read_unlock(); 665 goto xdp_xmit; 666 default: 667 bpf_warn_invalid_xdp_action(act); 668 case XDP_ABORTED: 669 trace_xdp_exception(vi->dev, xdp_prog, act); 670 case XDP_DROP: 671 goto err_xdp; 672 } 673 } 674 rcu_read_unlock(); 675 676 skb = build_skb(buf, buflen); 677 if (!skb) { 678 put_page(page); 679 goto err; 680 } 681 skb_reserve(skb, headroom - delta); 682 skb_put(skb, len); 683 if (!delta) { 684 buf += header_offset; 685 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len); 686 } /* keep zeroed vnet hdr since packet was changed by bpf */ 687 688 err: 689 return skb; 690 691 err_xdp: 692 rcu_read_unlock(); 693 dev->stats.rx_dropped++; 694 put_page(page); 695 xdp_xmit: 696 return NULL; 697 } 698 699 static struct sk_buff *receive_big(struct net_device *dev, 700 struct virtnet_info *vi, 701 struct receive_queue *rq, 702 void *buf, 703 unsigned int len) 704 { 705 struct page *page = buf; 706 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE); 707 708 if (unlikely(!skb)) 709 goto err; 710 711 return skb; 712 713 err: 714 dev->stats.rx_dropped++; 715 give_pages(rq, page); 716 return NULL; 717 } 718 719 static struct sk_buff *receive_mergeable(struct net_device *dev, 720 struct virtnet_info *vi, 721 struct receive_queue *rq, 722 void *buf, 723 void *ctx, 724 unsigned int len, 725 bool *xdp_xmit) 726 { 727 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 728 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 729 struct page *page = virt_to_head_page(buf); 730 int offset = buf - page_address(page); 731 struct sk_buff *head_skb, *curr_skb; 732 struct bpf_prog *xdp_prog; 733 unsigned int truesize; 734 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 735 int err; 736 737 head_skb = NULL; 738 739 rcu_read_lock(); 740 xdp_prog = rcu_dereference(rq->xdp_prog); 741 if (xdp_prog) { 742 struct xdp_frame *xdpf; 743 struct page *xdp_page; 744 struct xdp_buff xdp; 745 void *data; 746 u32 act; 747 748 /* Transient failure which in theory could occur if 749 * in-flight packets from before XDP was enabled reach 750 * the receive path after XDP is loaded. 751 */ 752 if (unlikely(hdr->hdr.gso_type)) 753 goto err_xdp; 754 755 /* This happens when rx buffer size is underestimated 756 * or headroom is not enough because of the buffer 757 * was refilled before XDP is set. This should only 758 * happen for the first several packets, so we don't 759 * care much about its performance. 760 */ 761 if (unlikely(num_buf > 1 || 762 headroom < virtnet_get_headroom(vi))) { 763 /* linearize data for XDP */ 764 xdp_page = xdp_linearize_page(rq, &num_buf, 765 page, offset, 766 VIRTIO_XDP_HEADROOM, 767 &len); 768 if (!xdp_page) 769 goto err_xdp; 770 offset = VIRTIO_XDP_HEADROOM; 771 } else { 772 xdp_page = page; 773 } 774 775 /* Allow consuming headroom but reserve enough space to push 776 * the descriptor on if we get an XDP_TX return code. 777 */ 778 data = page_address(xdp_page) + offset; 779 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len; 780 xdp.data = data + vi->hdr_len; 781 xdp_set_data_meta_invalid(&xdp); 782 xdp.data_end = xdp.data + (len - vi->hdr_len); 783 xdp.rxq = &rq->xdp_rxq; 784 785 act = bpf_prog_run_xdp(xdp_prog, &xdp); 786 787 switch (act) { 788 case XDP_PASS: 789 /* recalculate offset to account for any header 790 * adjustments. Note other cases do not build an 791 * skb and avoid using offset 792 */ 793 offset = xdp.data - 794 page_address(xdp_page) - vi->hdr_len; 795 796 /* recalculate len if xdp.data or xdp.data_end were 797 * adjusted 798 */ 799 len = xdp.data_end - xdp.data + vi->hdr_len; 800 /* We can only create skb based on xdp_page. */ 801 if (unlikely(xdp_page != page)) { 802 rcu_read_unlock(); 803 put_page(page); 804 head_skb = page_to_skb(vi, rq, xdp_page, 805 offset, len, PAGE_SIZE); 806 return head_skb; 807 } 808 break; 809 case XDP_TX: 810 xdpf = convert_to_xdp_frame(&xdp); 811 if (unlikely(!xdpf)) 812 goto err_xdp; 813 err = __virtnet_xdp_tx_xmit(vi, xdpf); 814 if (unlikely(err)) { 815 trace_xdp_exception(vi->dev, xdp_prog, act); 816 if (unlikely(xdp_page != page)) 817 put_page(xdp_page); 818 goto err_xdp; 819 } 820 *xdp_xmit = true; 821 if (unlikely(xdp_page != page)) 822 put_page(page); 823 rcu_read_unlock(); 824 goto xdp_xmit; 825 case XDP_REDIRECT: 826 err = xdp_do_redirect(dev, &xdp, xdp_prog); 827 if (err) { 828 if (unlikely(xdp_page != page)) 829 put_page(xdp_page); 830 goto err_xdp; 831 } 832 *xdp_xmit = true; 833 if (unlikely(xdp_page != page)) 834 put_page(page); 835 rcu_read_unlock(); 836 goto xdp_xmit; 837 default: 838 bpf_warn_invalid_xdp_action(act); 839 case XDP_ABORTED: 840 trace_xdp_exception(vi->dev, xdp_prog, act); 841 case XDP_DROP: 842 if (unlikely(xdp_page != page)) 843 __free_pages(xdp_page, 0); 844 goto err_xdp; 845 } 846 } 847 rcu_read_unlock(); 848 849 truesize = mergeable_ctx_to_truesize(ctx); 850 if (unlikely(len > truesize)) { 851 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 852 dev->name, len, (unsigned long)ctx); 853 dev->stats.rx_length_errors++; 854 goto err_skb; 855 } 856 857 head_skb = page_to_skb(vi, rq, page, offset, len, truesize); 858 curr_skb = head_skb; 859 860 if (unlikely(!curr_skb)) 861 goto err_skb; 862 while (--num_buf) { 863 int num_skb_frags; 864 865 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 866 if (unlikely(!buf)) { 867 pr_debug("%s: rx error: %d buffers out of %d missing\n", 868 dev->name, num_buf, 869 virtio16_to_cpu(vi->vdev, 870 hdr->num_buffers)); 871 dev->stats.rx_length_errors++; 872 goto err_buf; 873 } 874 875 page = virt_to_head_page(buf); 876 877 truesize = mergeable_ctx_to_truesize(ctx); 878 if (unlikely(len > truesize)) { 879 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 880 dev->name, len, (unsigned long)ctx); 881 dev->stats.rx_length_errors++; 882 goto err_skb; 883 } 884 885 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 886 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 887 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 888 889 if (unlikely(!nskb)) 890 goto err_skb; 891 if (curr_skb == head_skb) 892 skb_shinfo(curr_skb)->frag_list = nskb; 893 else 894 curr_skb->next = nskb; 895 curr_skb = nskb; 896 head_skb->truesize += nskb->truesize; 897 num_skb_frags = 0; 898 } 899 if (curr_skb != head_skb) { 900 head_skb->data_len += len; 901 head_skb->len += len; 902 head_skb->truesize += truesize; 903 } 904 offset = buf - page_address(page); 905 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 906 put_page(page); 907 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 908 len, truesize); 909 } else { 910 skb_add_rx_frag(curr_skb, num_skb_frags, page, 911 offset, len, truesize); 912 } 913 } 914 915 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 916 return head_skb; 917 918 err_xdp: 919 rcu_read_unlock(); 920 err_skb: 921 put_page(page); 922 while (num_buf-- > 1) { 923 buf = virtqueue_get_buf(rq->vq, &len); 924 if (unlikely(!buf)) { 925 pr_debug("%s: rx error: %d buffers missing\n", 926 dev->name, num_buf); 927 dev->stats.rx_length_errors++; 928 break; 929 } 930 page = virt_to_head_page(buf); 931 put_page(page); 932 } 933 err_buf: 934 dev->stats.rx_dropped++; 935 dev_kfree_skb(head_skb); 936 xdp_xmit: 937 return NULL; 938 } 939 940 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 941 void *buf, unsigned int len, void **ctx, bool *xdp_xmit) 942 { 943 struct net_device *dev = vi->dev; 944 struct sk_buff *skb; 945 struct virtio_net_hdr_mrg_rxbuf *hdr; 946 int ret; 947 948 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 949 pr_debug("%s: short packet %i\n", dev->name, len); 950 dev->stats.rx_length_errors++; 951 if (vi->mergeable_rx_bufs) { 952 put_page(virt_to_head_page(buf)); 953 } else if (vi->big_packets) { 954 give_pages(rq, buf); 955 } else { 956 put_page(virt_to_head_page(buf)); 957 } 958 return 0; 959 } 960 961 if (vi->mergeable_rx_bufs) 962 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit); 963 else if (vi->big_packets) 964 skb = receive_big(dev, vi, rq, buf, len); 965 else 966 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit); 967 968 if (unlikely(!skb)) 969 return 0; 970 971 hdr = skb_vnet_hdr(skb); 972 973 ret = skb->len; 974 975 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 976 skb->ip_summed = CHECKSUM_UNNECESSARY; 977 978 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 979 virtio_is_little_endian(vi->vdev))) { 980 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 981 dev->name, hdr->hdr.gso_type, 982 hdr->hdr.gso_size); 983 goto frame_err; 984 } 985 986 skb->protocol = eth_type_trans(skb, dev); 987 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 988 ntohs(skb->protocol), skb->len, skb->pkt_type); 989 990 napi_gro_receive(&rq->napi, skb); 991 return ret; 992 993 frame_err: 994 dev->stats.rx_frame_errors++; 995 dev_kfree_skb(skb); 996 return 0; 997 } 998 999 /* Unlike mergeable buffers, all buffers are allocated to the 1000 * same size, except for the headroom. For this reason we do 1001 * not need to use mergeable_len_to_ctx here - it is enough 1002 * to store the headroom as the context ignoring the truesize. 1003 */ 1004 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 1005 gfp_t gfp) 1006 { 1007 struct page_frag *alloc_frag = &rq->alloc_frag; 1008 char *buf; 1009 unsigned int xdp_headroom = virtnet_get_headroom(vi); 1010 void *ctx = (void *)(unsigned long)xdp_headroom; 1011 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 1012 int err; 1013 1014 len = SKB_DATA_ALIGN(len) + 1015 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1016 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 1017 return -ENOMEM; 1018 1019 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1020 get_page(alloc_frag->page); 1021 alloc_frag->offset += len; 1022 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 1023 vi->hdr_len + GOOD_PACKET_LEN); 1024 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1025 if (err < 0) 1026 put_page(virt_to_head_page(buf)); 1027 return err; 1028 } 1029 1030 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 1031 gfp_t gfp) 1032 { 1033 struct page *first, *list = NULL; 1034 char *p; 1035 int i, err, offset; 1036 1037 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2); 1038 1039 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */ 1040 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) { 1041 first = get_a_page(rq, gfp); 1042 if (!first) { 1043 if (list) 1044 give_pages(rq, list); 1045 return -ENOMEM; 1046 } 1047 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 1048 1049 /* chain new page in list head to match sg */ 1050 first->private = (unsigned long)list; 1051 list = first; 1052 } 1053 1054 first = get_a_page(rq, gfp); 1055 if (!first) { 1056 give_pages(rq, list); 1057 return -ENOMEM; 1058 } 1059 p = page_address(first); 1060 1061 /* rq->sg[0], rq->sg[1] share the same page */ 1062 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 1063 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 1064 1065 /* rq->sg[1] for data packet, from offset */ 1066 offset = sizeof(struct padded_vnet_hdr); 1067 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 1068 1069 /* chain first in list head */ 1070 first->private = (unsigned long)list; 1071 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2, 1072 first, gfp); 1073 if (err < 0) 1074 give_pages(rq, first); 1075 1076 return err; 1077 } 1078 1079 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 1080 struct ewma_pkt_len *avg_pkt_len, 1081 unsigned int room) 1082 { 1083 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 1084 unsigned int len; 1085 1086 if (room) 1087 return PAGE_SIZE - room; 1088 1089 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1090 rq->min_buf_len, PAGE_SIZE - hdr_len); 1091 1092 return ALIGN(len, L1_CACHE_BYTES); 1093 } 1094 1095 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1096 struct receive_queue *rq, gfp_t gfp) 1097 { 1098 struct page_frag *alloc_frag = &rq->alloc_frag; 1099 unsigned int headroom = virtnet_get_headroom(vi); 1100 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1101 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1102 char *buf; 1103 void *ctx; 1104 int err; 1105 unsigned int len, hole; 1106 1107 /* Extra tailroom is needed to satisfy XDP's assumption. This 1108 * means rx frags coalescing won't work, but consider we've 1109 * disabled GSO for XDP, it won't be a big issue. 1110 */ 1111 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room); 1112 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp))) 1113 return -ENOMEM; 1114 1115 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1116 buf += headroom; /* advance address leaving hole at front of pkt */ 1117 get_page(alloc_frag->page); 1118 alloc_frag->offset += len + room; 1119 hole = alloc_frag->size - alloc_frag->offset; 1120 if (hole < len + room) { 1121 /* To avoid internal fragmentation, if there is very likely not 1122 * enough space for another buffer, add the remaining space to 1123 * the current buffer. 1124 */ 1125 len += hole; 1126 alloc_frag->offset += hole; 1127 } 1128 1129 sg_init_one(rq->sg, buf, len); 1130 ctx = mergeable_len_to_ctx(len, headroom); 1131 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1132 if (err < 0) 1133 put_page(virt_to_head_page(buf)); 1134 1135 return err; 1136 } 1137 1138 /* 1139 * Returns false if we couldn't fill entirely (OOM). 1140 * 1141 * Normally run in the receive path, but can also be run from ndo_open 1142 * before we're receiving packets, or from refill_work which is 1143 * careful to disable receiving (using napi_disable). 1144 */ 1145 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1146 gfp_t gfp) 1147 { 1148 int err; 1149 bool oom; 1150 1151 do { 1152 if (vi->mergeable_rx_bufs) 1153 err = add_recvbuf_mergeable(vi, rq, gfp); 1154 else if (vi->big_packets) 1155 err = add_recvbuf_big(vi, rq, gfp); 1156 else 1157 err = add_recvbuf_small(vi, rq, gfp); 1158 1159 oom = err == -ENOMEM; 1160 if (err) 1161 break; 1162 } while (rq->vq->num_free); 1163 virtqueue_kick(rq->vq); 1164 return !oom; 1165 } 1166 1167 static void skb_recv_done(struct virtqueue *rvq) 1168 { 1169 struct virtnet_info *vi = rvq->vdev->priv; 1170 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1171 1172 virtqueue_napi_schedule(&rq->napi, rvq); 1173 } 1174 1175 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1176 { 1177 napi_enable(napi); 1178 1179 /* If all buffers were filled by other side before we napi_enabled, we 1180 * won't get another interrupt, so process any outstanding packets now. 1181 * Call local_bh_enable after to trigger softIRQ processing. 1182 */ 1183 local_bh_disable(); 1184 virtqueue_napi_schedule(napi, vq); 1185 local_bh_enable(); 1186 } 1187 1188 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1189 struct virtqueue *vq, 1190 struct napi_struct *napi) 1191 { 1192 if (!napi->weight) 1193 return; 1194 1195 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1196 * enable the feature if this is likely affine with the transmit path. 1197 */ 1198 if (!vi->affinity_hint_set) { 1199 napi->weight = 0; 1200 return; 1201 } 1202 1203 return virtnet_napi_enable(vq, napi); 1204 } 1205 1206 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1207 { 1208 if (napi->weight) 1209 napi_disable(napi); 1210 } 1211 1212 static void refill_work(struct work_struct *work) 1213 { 1214 struct virtnet_info *vi = 1215 container_of(work, struct virtnet_info, refill.work); 1216 bool still_empty; 1217 int i; 1218 1219 for (i = 0; i < vi->curr_queue_pairs; i++) { 1220 struct receive_queue *rq = &vi->rq[i]; 1221 1222 napi_disable(&rq->napi); 1223 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1224 virtnet_napi_enable(rq->vq, &rq->napi); 1225 1226 /* In theory, this can happen: if we don't get any buffers in 1227 * we will *never* try to fill again. 1228 */ 1229 if (still_empty) 1230 schedule_delayed_work(&vi->refill, HZ/2); 1231 } 1232 } 1233 1234 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit) 1235 { 1236 struct virtnet_info *vi = rq->vq->vdev->priv; 1237 unsigned int len, received = 0, bytes = 0; 1238 void *buf; 1239 1240 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1241 void *ctx; 1242 1243 while (received < budget && 1244 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1245 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit); 1246 received++; 1247 } 1248 } else { 1249 while (received < budget && 1250 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1251 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit); 1252 received++; 1253 } 1254 } 1255 1256 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) { 1257 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) 1258 schedule_delayed_work(&vi->refill, 0); 1259 } 1260 1261 u64_stats_update_begin(&rq->stats.syncp); 1262 rq->stats.bytes += bytes; 1263 rq->stats.packets += received; 1264 u64_stats_update_end(&rq->stats.syncp); 1265 1266 return received; 1267 } 1268 1269 static void free_old_xmit_skbs(struct send_queue *sq) 1270 { 1271 struct sk_buff *skb; 1272 unsigned int len; 1273 unsigned int packets = 0; 1274 unsigned int bytes = 0; 1275 1276 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) { 1277 pr_debug("Sent skb %p\n", skb); 1278 1279 bytes += skb->len; 1280 packets++; 1281 1282 dev_consume_skb_any(skb); 1283 } 1284 1285 /* Avoid overhead when no packets have been processed 1286 * happens when called speculatively from start_xmit. 1287 */ 1288 if (!packets) 1289 return; 1290 1291 u64_stats_update_begin(&sq->stats.syncp); 1292 sq->stats.bytes += bytes; 1293 sq->stats.packets += packets; 1294 u64_stats_update_end(&sq->stats.syncp); 1295 } 1296 1297 static void virtnet_poll_cleantx(struct receive_queue *rq) 1298 { 1299 struct virtnet_info *vi = rq->vq->vdev->priv; 1300 unsigned int index = vq2rxq(rq->vq); 1301 struct send_queue *sq = &vi->sq[index]; 1302 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1303 1304 if (!sq->napi.weight) 1305 return; 1306 1307 if (__netif_tx_trylock(txq)) { 1308 free_old_xmit_skbs(sq); 1309 __netif_tx_unlock(txq); 1310 } 1311 1312 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1313 netif_tx_wake_queue(txq); 1314 } 1315 1316 static int virtnet_poll(struct napi_struct *napi, int budget) 1317 { 1318 struct receive_queue *rq = 1319 container_of(napi, struct receive_queue, napi); 1320 struct virtnet_info *vi = rq->vq->vdev->priv; 1321 struct send_queue *sq; 1322 unsigned int received, qp; 1323 bool xdp_xmit = false; 1324 1325 virtnet_poll_cleantx(rq); 1326 1327 received = virtnet_receive(rq, budget, &xdp_xmit); 1328 1329 /* Out of packets? */ 1330 if (received < budget) 1331 virtqueue_napi_complete(napi, rq->vq, received); 1332 1333 if (xdp_xmit) { 1334 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + 1335 smp_processor_id(); 1336 sq = &vi->sq[qp]; 1337 virtqueue_kick(sq->vq); 1338 xdp_do_flush_map(); 1339 } 1340 1341 return received; 1342 } 1343 1344 static int virtnet_open(struct net_device *dev) 1345 { 1346 struct virtnet_info *vi = netdev_priv(dev); 1347 int i, err; 1348 1349 for (i = 0; i < vi->max_queue_pairs; i++) { 1350 if (i < vi->curr_queue_pairs) 1351 /* Make sure we have some buffers: if oom use wq. */ 1352 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1353 schedule_delayed_work(&vi->refill, 0); 1354 1355 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i); 1356 if (err < 0) 1357 return err; 1358 1359 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq, 1360 MEM_TYPE_PAGE_SHARED, NULL); 1361 if (err < 0) { 1362 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1363 return err; 1364 } 1365 1366 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 1367 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 1368 } 1369 1370 return 0; 1371 } 1372 1373 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 1374 { 1375 struct send_queue *sq = container_of(napi, struct send_queue, napi); 1376 struct virtnet_info *vi = sq->vq->vdev->priv; 1377 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq)); 1378 1379 __netif_tx_lock(txq, raw_smp_processor_id()); 1380 free_old_xmit_skbs(sq); 1381 __netif_tx_unlock(txq); 1382 1383 virtqueue_napi_complete(napi, sq->vq, 0); 1384 1385 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1386 netif_tx_wake_queue(txq); 1387 1388 return 0; 1389 } 1390 1391 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 1392 { 1393 struct virtio_net_hdr_mrg_rxbuf *hdr; 1394 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 1395 struct virtnet_info *vi = sq->vq->vdev->priv; 1396 int num_sg; 1397 unsigned hdr_len = vi->hdr_len; 1398 bool can_push; 1399 1400 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 1401 1402 can_push = vi->any_header_sg && 1403 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 1404 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 1405 /* Even if we can, don't push here yet as this would skew 1406 * csum_start offset below. */ 1407 if (can_push) 1408 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 1409 else 1410 hdr = skb_vnet_hdr(skb); 1411 1412 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 1413 virtio_is_little_endian(vi->vdev), false)) 1414 BUG(); 1415 1416 if (vi->mergeable_rx_bufs) 1417 hdr->num_buffers = 0; 1418 1419 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 1420 if (can_push) { 1421 __skb_push(skb, hdr_len); 1422 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 1423 if (unlikely(num_sg < 0)) 1424 return num_sg; 1425 /* Pull header back to avoid skew in tx bytes calculations. */ 1426 __skb_pull(skb, hdr_len); 1427 } else { 1428 sg_set_buf(sq->sg, hdr, hdr_len); 1429 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 1430 if (unlikely(num_sg < 0)) 1431 return num_sg; 1432 num_sg++; 1433 } 1434 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 1435 } 1436 1437 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 1438 { 1439 struct virtnet_info *vi = netdev_priv(dev); 1440 int qnum = skb_get_queue_mapping(skb); 1441 struct send_queue *sq = &vi->sq[qnum]; 1442 int err; 1443 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 1444 bool kick = !skb->xmit_more; 1445 bool use_napi = sq->napi.weight; 1446 1447 /* Free up any pending old buffers before queueing new ones. */ 1448 free_old_xmit_skbs(sq); 1449 1450 if (use_napi && kick) 1451 virtqueue_enable_cb_delayed(sq->vq); 1452 1453 /* timestamp packet in software */ 1454 skb_tx_timestamp(skb); 1455 1456 /* Try to transmit */ 1457 err = xmit_skb(sq, skb); 1458 1459 /* This should not happen! */ 1460 if (unlikely(err)) { 1461 dev->stats.tx_fifo_errors++; 1462 if (net_ratelimit()) 1463 dev_warn(&dev->dev, 1464 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err); 1465 dev->stats.tx_dropped++; 1466 dev_kfree_skb_any(skb); 1467 return NETDEV_TX_OK; 1468 } 1469 1470 /* Don't wait up for transmitted skbs to be freed. */ 1471 if (!use_napi) { 1472 skb_orphan(skb); 1473 nf_reset(skb); 1474 } 1475 1476 /* If running out of space, stop queue to avoid getting packets that we 1477 * are then unable to transmit. 1478 * An alternative would be to force queuing layer to requeue the skb by 1479 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be 1480 * returned in a normal path of operation: it means that driver is not 1481 * maintaining the TX queue stop/start state properly, and causes 1482 * the stack to do a non-trivial amount of useless work. 1483 * Since most packets only take 1 or 2 ring slots, stopping the queue 1484 * early means 16 slots are typically wasted. 1485 */ 1486 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { 1487 netif_stop_subqueue(dev, qnum); 1488 if (!use_napi && 1489 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { 1490 /* More just got used, free them then recheck. */ 1491 free_old_xmit_skbs(sq); 1492 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { 1493 netif_start_subqueue(dev, qnum); 1494 virtqueue_disable_cb(sq->vq); 1495 } 1496 } 1497 } 1498 1499 if (kick || netif_xmit_stopped(txq)) 1500 virtqueue_kick(sq->vq); 1501 1502 return NETDEV_TX_OK; 1503 } 1504 1505 /* 1506 * Send command via the control virtqueue and check status. Commands 1507 * supported by the hypervisor, as indicated by feature bits, should 1508 * never fail unless improperly formatted. 1509 */ 1510 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 1511 struct scatterlist *out) 1512 { 1513 struct scatterlist *sgs[4], hdr, stat; 1514 unsigned out_num = 0, tmp; 1515 1516 /* Caller should know better */ 1517 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 1518 1519 vi->ctrl->status = ~0; 1520 vi->ctrl->hdr.class = class; 1521 vi->ctrl->hdr.cmd = cmd; 1522 /* Add header */ 1523 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr)); 1524 sgs[out_num++] = &hdr; 1525 1526 if (out) 1527 sgs[out_num++] = out; 1528 1529 /* Add return status. */ 1530 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status)); 1531 sgs[out_num] = &stat; 1532 1533 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 1534 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 1535 1536 if (unlikely(!virtqueue_kick(vi->cvq))) 1537 return vi->ctrl->status == VIRTIO_NET_OK; 1538 1539 /* Spin for a response, the kick causes an ioport write, trapping 1540 * into the hypervisor, so the request should be handled immediately. 1541 */ 1542 while (!virtqueue_get_buf(vi->cvq, &tmp) && 1543 !virtqueue_is_broken(vi->cvq)) 1544 cpu_relax(); 1545 1546 return vi->ctrl->status == VIRTIO_NET_OK; 1547 } 1548 1549 static int virtnet_set_mac_address(struct net_device *dev, void *p) 1550 { 1551 struct virtnet_info *vi = netdev_priv(dev); 1552 struct virtio_device *vdev = vi->vdev; 1553 int ret; 1554 struct sockaddr *addr; 1555 struct scatterlist sg; 1556 1557 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 1558 if (!addr) 1559 return -ENOMEM; 1560 1561 ret = eth_prepare_mac_addr_change(dev, addr); 1562 if (ret) 1563 goto out; 1564 1565 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 1566 sg_init_one(&sg, addr->sa_data, dev->addr_len); 1567 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1568 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 1569 dev_warn(&vdev->dev, 1570 "Failed to set mac address by vq command.\n"); 1571 ret = -EINVAL; 1572 goto out; 1573 } 1574 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 1575 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 1576 unsigned int i; 1577 1578 /* Naturally, this has an atomicity problem. */ 1579 for (i = 0; i < dev->addr_len; i++) 1580 virtio_cwrite8(vdev, 1581 offsetof(struct virtio_net_config, mac) + 1582 i, addr->sa_data[i]); 1583 } 1584 1585 eth_commit_mac_addr_change(dev, p); 1586 ret = 0; 1587 1588 out: 1589 kfree(addr); 1590 return ret; 1591 } 1592 1593 static void virtnet_stats(struct net_device *dev, 1594 struct rtnl_link_stats64 *tot) 1595 { 1596 struct virtnet_info *vi = netdev_priv(dev); 1597 unsigned int start; 1598 int i; 1599 1600 for (i = 0; i < vi->max_queue_pairs; i++) { 1601 u64 tpackets, tbytes, rpackets, rbytes; 1602 struct receive_queue *rq = &vi->rq[i]; 1603 struct send_queue *sq = &vi->sq[i]; 1604 1605 do { 1606 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1607 tpackets = sq->stats.packets; 1608 tbytes = sq->stats.bytes; 1609 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1610 1611 do { 1612 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1613 rpackets = rq->stats.packets; 1614 rbytes = rq->stats.bytes; 1615 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 1616 1617 tot->rx_packets += rpackets; 1618 tot->tx_packets += tpackets; 1619 tot->rx_bytes += rbytes; 1620 tot->tx_bytes += tbytes; 1621 } 1622 1623 tot->tx_dropped = dev->stats.tx_dropped; 1624 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 1625 tot->rx_dropped = dev->stats.rx_dropped; 1626 tot->rx_length_errors = dev->stats.rx_length_errors; 1627 tot->rx_frame_errors = dev->stats.rx_frame_errors; 1628 } 1629 1630 #ifdef CONFIG_NET_POLL_CONTROLLER 1631 static void virtnet_netpoll(struct net_device *dev) 1632 { 1633 struct virtnet_info *vi = netdev_priv(dev); 1634 int i; 1635 1636 for (i = 0; i < vi->curr_queue_pairs; i++) 1637 napi_schedule(&vi->rq[i].napi); 1638 } 1639 #endif 1640 1641 static void virtnet_ack_link_announce(struct virtnet_info *vi) 1642 { 1643 rtnl_lock(); 1644 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 1645 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 1646 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 1647 rtnl_unlock(); 1648 } 1649 1650 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1651 { 1652 struct scatterlist sg; 1653 struct net_device *dev = vi->dev; 1654 1655 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 1656 return 0; 1657 1658 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 1659 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq)); 1660 1661 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 1662 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 1663 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 1664 queue_pairs); 1665 return -EINVAL; 1666 } else { 1667 vi->curr_queue_pairs = queue_pairs; 1668 /* virtnet_open() will refill when device is going to up. */ 1669 if (dev->flags & IFF_UP) 1670 schedule_delayed_work(&vi->refill, 0); 1671 } 1672 1673 return 0; 1674 } 1675 1676 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1677 { 1678 int err; 1679 1680 rtnl_lock(); 1681 err = _virtnet_set_queues(vi, queue_pairs); 1682 rtnl_unlock(); 1683 return err; 1684 } 1685 1686 static int virtnet_close(struct net_device *dev) 1687 { 1688 struct virtnet_info *vi = netdev_priv(dev); 1689 int i; 1690 1691 /* Make sure refill_work doesn't re-enable napi! */ 1692 cancel_delayed_work_sync(&vi->refill); 1693 1694 for (i = 0; i < vi->max_queue_pairs; i++) { 1695 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1696 napi_disable(&vi->rq[i].napi); 1697 virtnet_napi_tx_disable(&vi->sq[i].napi); 1698 } 1699 1700 return 0; 1701 } 1702 1703 static void virtnet_set_rx_mode(struct net_device *dev) 1704 { 1705 struct virtnet_info *vi = netdev_priv(dev); 1706 struct scatterlist sg[2]; 1707 struct virtio_net_ctrl_mac *mac_data; 1708 struct netdev_hw_addr *ha; 1709 int uc_count; 1710 int mc_count; 1711 void *buf; 1712 int i; 1713 1714 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 1715 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 1716 return; 1717 1718 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0); 1719 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 1720 1721 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc)); 1722 1723 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1724 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 1725 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 1726 vi->ctrl->promisc ? "en" : "dis"); 1727 1728 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti)); 1729 1730 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1731 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 1732 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 1733 vi->ctrl->allmulti ? "en" : "dis"); 1734 1735 uc_count = netdev_uc_count(dev); 1736 mc_count = netdev_mc_count(dev); 1737 /* MAC filter - use one buffer for both lists */ 1738 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 1739 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 1740 mac_data = buf; 1741 if (!buf) 1742 return; 1743 1744 sg_init_table(sg, 2); 1745 1746 /* Store the unicast list and count in the front of the buffer */ 1747 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 1748 i = 0; 1749 netdev_for_each_uc_addr(ha, dev) 1750 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1751 1752 sg_set_buf(&sg[0], mac_data, 1753 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 1754 1755 /* multicast list and count fill the end */ 1756 mac_data = (void *)&mac_data->macs[uc_count][0]; 1757 1758 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 1759 i = 0; 1760 netdev_for_each_mc_addr(ha, dev) 1761 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1762 1763 sg_set_buf(&sg[1], mac_data, 1764 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 1765 1766 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1767 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 1768 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 1769 1770 kfree(buf); 1771 } 1772 1773 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 1774 __be16 proto, u16 vid) 1775 { 1776 struct virtnet_info *vi = netdev_priv(dev); 1777 struct scatterlist sg; 1778 1779 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 1780 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 1781 1782 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1783 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 1784 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 1785 return 0; 1786 } 1787 1788 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 1789 __be16 proto, u16 vid) 1790 { 1791 struct virtnet_info *vi = netdev_priv(dev); 1792 struct scatterlist sg; 1793 1794 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 1795 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 1796 1797 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1798 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 1799 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 1800 return 0; 1801 } 1802 1803 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu) 1804 { 1805 int i; 1806 1807 if (vi->affinity_hint_set) { 1808 for (i = 0; i < vi->max_queue_pairs; i++) { 1809 virtqueue_set_affinity(vi->rq[i].vq, -1); 1810 virtqueue_set_affinity(vi->sq[i].vq, -1); 1811 } 1812 1813 vi->affinity_hint_set = false; 1814 } 1815 } 1816 1817 static void virtnet_set_affinity(struct virtnet_info *vi) 1818 { 1819 int i; 1820 int cpu; 1821 1822 /* In multiqueue mode, when the number of cpu is equal to the number of 1823 * queue pairs, we let the queue pairs to be private to one cpu by 1824 * setting the affinity hint to eliminate the contention. 1825 */ 1826 if (vi->curr_queue_pairs == 1 || 1827 vi->max_queue_pairs != num_online_cpus()) { 1828 virtnet_clean_affinity(vi, -1); 1829 return; 1830 } 1831 1832 i = 0; 1833 for_each_online_cpu(cpu) { 1834 virtqueue_set_affinity(vi->rq[i].vq, cpu); 1835 virtqueue_set_affinity(vi->sq[i].vq, cpu); 1836 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i); 1837 i++; 1838 } 1839 1840 vi->affinity_hint_set = true; 1841 } 1842 1843 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 1844 { 1845 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1846 node); 1847 virtnet_set_affinity(vi); 1848 return 0; 1849 } 1850 1851 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 1852 { 1853 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1854 node_dead); 1855 virtnet_set_affinity(vi); 1856 return 0; 1857 } 1858 1859 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 1860 { 1861 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1862 node); 1863 1864 virtnet_clean_affinity(vi, cpu); 1865 return 0; 1866 } 1867 1868 static enum cpuhp_state virtionet_online; 1869 1870 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 1871 { 1872 int ret; 1873 1874 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 1875 if (ret) 1876 return ret; 1877 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1878 &vi->node_dead); 1879 if (!ret) 1880 return ret; 1881 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1882 return ret; 1883 } 1884 1885 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 1886 { 1887 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1888 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1889 &vi->node_dead); 1890 } 1891 1892 static void virtnet_get_ringparam(struct net_device *dev, 1893 struct ethtool_ringparam *ring) 1894 { 1895 struct virtnet_info *vi = netdev_priv(dev); 1896 1897 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq); 1898 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq); 1899 ring->rx_pending = ring->rx_max_pending; 1900 ring->tx_pending = ring->tx_max_pending; 1901 } 1902 1903 1904 static void virtnet_get_drvinfo(struct net_device *dev, 1905 struct ethtool_drvinfo *info) 1906 { 1907 struct virtnet_info *vi = netdev_priv(dev); 1908 struct virtio_device *vdev = vi->vdev; 1909 1910 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 1911 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 1912 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 1913 1914 } 1915 1916 /* TODO: Eliminate OOO packets during switching */ 1917 static int virtnet_set_channels(struct net_device *dev, 1918 struct ethtool_channels *channels) 1919 { 1920 struct virtnet_info *vi = netdev_priv(dev); 1921 u16 queue_pairs = channels->combined_count; 1922 int err; 1923 1924 /* We don't support separate rx/tx channels. 1925 * We don't allow setting 'other' channels. 1926 */ 1927 if (channels->rx_count || channels->tx_count || channels->other_count) 1928 return -EINVAL; 1929 1930 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 1931 return -EINVAL; 1932 1933 /* For now we don't support modifying channels while XDP is loaded 1934 * also when XDP is loaded all RX queues have XDP programs so we only 1935 * need to check a single RX queue. 1936 */ 1937 if (vi->rq[0].xdp_prog) 1938 return -EINVAL; 1939 1940 get_online_cpus(); 1941 err = _virtnet_set_queues(vi, queue_pairs); 1942 if (!err) { 1943 netif_set_real_num_tx_queues(dev, queue_pairs); 1944 netif_set_real_num_rx_queues(dev, queue_pairs); 1945 1946 virtnet_set_affinity(vi); 1947 } 1948 put_online_cpus(); 1949 1950 return err; 1951 } 1952 1953 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 1954 { 1955 struct virtnet_info *vi = netdev_priv(dev); 1956 char *p = (char *)data; 1957 unsigned int i, j; 1958 1959 switch (stringset) { 1960 case ETH_SS_STATS: 1961 for (i = 0; i < vi->curr_queue_pairs; i++) { 1962 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 1963 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s", 1964 i, virtnet_rq_stats_desc[j].desc); 1965 p += ETH_GSTRING_LEN; 1966 } 1967 } 1968 1969 for (i = 0; i < vi->curr_queue_pairs; i++) { 1970 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 1971 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s", 1972 i, virtnet_sq_stats_desc[j].desc); 1973 p += ETH_GSTRING_LEN; 1974 } 1975 } 1976 break; 1977 } 1978 } 1979 1980 static int virtnet_get_sset_count(struct net_device *dev, int sset) 1981 { 1982 struct virtnet_info *vi = netdev_priv(dev); 1983 1984 switch (sset) { 1985 case ETH_SS_STATS: 1986 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 1987 VIRTNET_SQ_STATS_LEN); 1988 default: 1989 return -EOPNOTSUPP; 1990 } 1991 } 1992 1993 static void virtnet_get_ethtool_stats(struct net_device *dev, 1994 struct ethtool_stats *stats, u64 *data) 1995 { 1996 struct virtnet_info *vi = netdev_priv(dev); 1997 unsigned int idx = 0, start, i, j; 1998 const u8 *stats_base; 1999 size_t offset; 2000 2001 for (i = 0; i < vi->curr_queue_pairs; i++) { 2002 struct receive_queue *rq = &vi->rq[i]; 2003 2004 stats_base = (u8 *)&rq->stats; 2005 do { 2006 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 2007 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 2008 offset = virtnet_rq_stats_desc[j].offset; 2009 data[idx + j] = *(u64 *)(stats_base + offset); 2010 } 2011 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 2012 idx += VIRTNET_RQ_STATS_LEN; 2013 } 2014 2015 for (i = 0; i < vi->curr_queue_pairs; i++) { 2016 struct send_queue *sq = &vi->sq[i]; 2017 2018 stats_base = (u8 *)&sq->stats; 2019 do { 2020 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 2021 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 2022 offset = virtnet_sq_stats_desc[j].offset; 2023 data[idx + j] = *(u64 *)(stats_base + offset); 2024 } 2025 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 2026 idx += VIRTNET_SQ_STATS_LEN; 2027 } 2028 } 2029 2030 static void virtnet_get_channels(struct net_device *dev, 2031 struct ethtool_channels *channels) 2032 { 2033 struct virtnet_info *vi = netdev_priv(dev); 2034 2035 channels->combined_count = vi->curr_queue_pairs; 2036 channels->max_combined = vi->max_queue_pairs; 2037 channels->max_other = 0; 2038 channels->rx_count = 0; 2039 channels->tx_count = 0; 2040 channels->other_count = 0; 2041 } 2042 2043 /* Check if the user is trying to change anything besides speed/duplex */ 2044 static bool 2045 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd) 2046 { 2047 struct ethtool_link_ksettings diff1 = *cmd; 2048 struct ethtool_link_ksettings diff2 = {}; 2049 2050 /* cmd is always set so we need to clear it, validate the port type 2051 * and also without autonegotiation we can ignore advertising 2052 */ 2053 diff1.base.speed = 0; 2054 diff2.base.port = PORT_OTHER; 2055 ethtool_link_ksettings_zero_link_mode(&diff1, advertising); 2056 diff1.base.duplex = 0; 2057 diff1.base.cmd = 0; 2058 diff1.base.link_mode_masks_nwords = 0; 2059 2060 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) && 2061 bitmap_empty(diff1.link_modes.supported, 2062 __ETHTOOL_LINK_MODE_MASK_NBITS) && 2063 bitmap_empty(diff1.link_modes.advertising, 2064 __ETHTOOL_LINK_MODE_MASK_NBITS) && 2065 bitmap_empty(diff1.link_modes.lp_advertising, 2066 __ETHTOOL_LINK_MODE_MASK_NBITS); 2067 } 2068 2069 static int virtnet_set_link_ksettings(struct net_device *dev, 2070 const struct ethtool_link_ksettings *cmd) 2071 { 2072 struct virtnet_info *vi = netdev_priv(dev); 2073 u32 speed; 2074 2075 speed = cmd->base.speed; 2076 /* don't allow custom speed and duplex */ 2077 if (!ethtool_validate_speed(speed) || 2078 !ethtool_validate_duplex(cmd->base.duplex) || 2079 !virtnet_validate_ethtool_cmd(cmd)) 2080 return -EINVAL; 2081 vi->speed = speed; 2082 vi->duplex = cmd->base.duplex; 2083 2084 return 0; 2085 } 2086 2087 static int virtnet_get_link_ksettings(struct net_device *dev, 2088 struct ethtool_link_ksettings *cmd) 2089 { 2090 struct virtnet_info *vi = netdev_priv(dev); 2091 2092 cmd->base.speed = vi->speed; 2093 cmd->base.duplex = vi->duplex; 2094 cmd->base.port = PORT_OTHER; 2095 2096 return 0; 2097 } 2098 2099 static void virtnet_init_settings(struct net_device *dev) 2100 { 2101 struct virtnet_info *vi = netdev_priv(dev); 2102 2103 vi->speed = SPEED_UNKNOWN; 2104 vi->duplex = DUPLEX_UNKNOWN; 2105 } 2106 2107 static void virtnet_update_settings(struct virtnet_info *vi) 2108 { 2109 u32 speed; 2110 u8 duplex; 2111 2112 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 2113 return; 2114 2115 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config, 2116 speed)); 2117 if (ethtool_validate_speed(speed)) 2118 vi->speed = speed; 2119 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config, 2120 duplex)); 2121 if (ethtool_validate_duplex(duplex)) 2122 vi->duplex = duplex; 2123 } 2124 2125 static const struct ethtool_ops virtnet_ethtool_ops = { 2126 .get_drvinfo = virtnet_get_drvinfo, 2127 .get_link = ethtool_op_get_link, 2128 .get_ringparam = virtnet_get_ringparam, 2129 .get_strings = virtnet_get_strings, 2130 .get_sset_count = virtnet_get_sset_count, 2131 .get_ethtool_stats = virtnet_get_ethtool_stats, 2132 .set_channels = virtnet_set_channels, 2133 .get_channels = virtnet_get_channels, 2134 .get_ts_info = ethtool_op_get_ts_info, 2135 .get_link_ksettings = virtnet_get_link_ksettings, 2136 .set_link_ksettings = virtnet_set_link_ksettings, 2137 }; 2138 2139 static void virtnet_freeze_down(struct virtio_device *vdev) 2140 { 2141 struct virtnet_info *vi = vdev->priv; 2142 int i; 2143 2144 /* Make sure no work handler is accessing the device */ 2145 flush_work(&vi->config_work); 2146 2147 netif_device_detach(vi->dev); 2148 netif_tx_disable(vi->dev); 2149 cancel_delayed_work_sync(&vi->refill); 2150 2151 if (netif_running(vi->dev)) { 2152 for (i = 0; i < vi->max_queue_pairs; i++) { 2153 napi_disable(&vi->rq[i].napi); 2154 virtnet_napi_tx_disable(&vi->sq[i].napi); 2155 } 2156 } 2157 } 2158 2159 static int init_vqs(struct virtnet_info *vi); 2160 2161 static int virtnet_restore_up(struct virtio_device *vdev) 2162 { 2163 struct virtnet_info *vi = vdev->priv; 2164 int err, i; 2165 2166 err = init_vqs(vi); 2167 if (err) 2168 return err; 2169 2170 virtio_device_ready(vdev); 2171 2172 if (netif_running(vi->dev)) { 2173 for (i = 0; i < vi->curr_queue_pairs; i++) 2174 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 2175 schedule_delayed_work(&vi->refill, 0); 2176 2177 for (i = 0; i < vi->max_queue_pairs; i++) { 2178 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2179 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 2180 &vi->sq[i].napi); 2181 } 2182 } 2183 2184 netif_device_attach(vi->dev); 2185 return err; 2186 } 2187 2188 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 2189 { 2190 struct scatterlist sg; 2191 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads); 2192 2193 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads)); 2194 2195 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 2196 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 2197 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n"); 2198 return -EINVAL; 2199 } 2200 2201 return 0; 2202 } 2203 2204 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 2205 { 2206 u64 offloads = 0; 2207 2208 if (!vi->guest_offloads) 2209 return 0; 2210 2211 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2212 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2213 2214 return virtnet_set_guest_offloads(vi, offloads); 2215 } 2216 2217 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 2218 { 2219 u64 offloads = vi->guest_offloads; 2220 2221 if (!vi->guest_offloads) 2222 return 0; 2223 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2224 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2225 2226 return virtnet_set_guest_offloads(vi, offloads); 2227 } 2228 2229 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 2230 struct netlink_ext_ack *extack) 2231 { 2232 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr); 2233 struct virtnet_info *vi = netdev_priv(dev); 2234 struct bpf_prog *old_prog; 2235 u16 xdp_qp = 0, curr_qp; 2236 int i, err; 2237 2238 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 2239 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 2240 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 2241 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 2242 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) { 2243 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first"); 2244 return -EOPNOTSUPP; 2245 } 2246 2247 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 2248 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 2249 return -EINVAL; 2250 } 2251 2252 if (dev->mtu > max_sz) { 2253 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP"); 2254 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz); 2255 return -EINVAL; 2256 } 2257 2258 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 2259 if (prog) 2260 xdp_qp = nr_cpu_ids; 2261 2262 /* XDP requires extra queues for XDP_TX */ 2263 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 2264 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available"); 2265 netdev_warn(dev, "request %i queues but max is %i\n", 2266 curr_qp + xdp_qp, vi->max_queue_pairs); 2267 return -ENOMEM; 2268 } 2269 2270 if (prog) { 2271 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1); 2272 if (IS_ERR(prog)) 2273 return PTR_ERR(prog); 2274 } 2275 2276 /* Make sure NAPI is not using any XDP TX queues for RX. */ 2277 if (netif_running(dev)) 2278 for (i = 0; i < vi->max_queue_pairs; i++) 2279 napi_disable(&vi->rq[i].napi); 2280 2281 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 2282 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 2283 if (err) 2284 goto err; 2285 vi->xdp_queue_pairs = xdp_qp; 2286 2287 for (i = 0; i < vi->max_queue_pairs; i++) { 2288 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2289 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 2290 if (i == 0) { 2291 if (!old_prog) 2292 virtnet_clear_guest_offloads(vi); 2293 if (!prog) 2294 virtnet_restore_guest_offloads(vi); 2295 } 2296 if (old_prog) 2297 bpf_prog_put(old_prog); 2298 if (netif_running(dev)) 2299 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2300 } 2301 2302 return 0; 2303 2304 err: 2305 for (i = 0; i < vi->max_queue_pairs; i++) 2306 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2307 if (prog) 2308 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 2309 return err; 2310 } 2311 2312 static u32 virtnet_xdp_query(struct net_device *dev) 2313 { 2314 struct virtnet_info *vi = netdev_priv(dev); 2315 const struct bpf_prog *xdp_prog; 2316 int i; 2317 2318 for (i = 0; i < vi->max_queue_pairs; i++) { 2319 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2320 if (xdp_prog) 2321 return xdp_prog->aux->id; 2322 } 2323 return 0; 2324 } 2325 2326 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 2327 { 2328 switch (xdp->command) { 2329 case XDP_SETUP_PROG: 2330 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 2331 case XDP_QUERY_PROG: 2332 xdp->prog_id = virtnet_xdp_query(dev); 2333 xdp->prog_attached = !!xdp->prog_id; 2334 return 0; 2335 default: 2336 return -EINVAL; 2337 } 2338 } 2339 2340 static const struct net_device_ops virtnet_netdev = { 2341 .ndo_open = virtnet_open, 2342 .ndo_stop = virtnet_close, 2343 .ndo_start_xmit = start_xmit, 2344 .ndo_validate_addr = eth_validate_addr, 2345 .ndo_set_mac_address = virtnet_set_mac_address, 2346 .ndo_set_rx_mode = virtnet_set_rx_mode, 2347 .ndo_get_stats64 = virtnet_stats, 2348 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 2349 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 2350 #ifdef CONFIG_NET_POLL_CONTROLLER 2351 .ndo_poll_controller = virtnet_netpoll, 2352 #endif 2353 .ndo_bpf = virtnet_xdp, 2354 .ndo_xdp_xmit = virtnet_xdp_xmit, 2355 .ndo_xdp_flush = virtnet_xdp_flush, 2356 .ndo_features_check = passthru_features_check, 2357 }; 2358 2359 static void virtnet_config_changed_work(struct work_struct *work) 2360 { 2361 struct virtnet_info *vi = 2362 container_of(work, struct virtnet_info, config_work); 2363 u16 v; 2364 2365 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 2366 struct virtio_net_config, status, &v) < 0) 2367 return; 2368 2369 if (v & VIRTIO_NET_S_ANNOUNCE) { 2370 netdev_notify_peers(vi->dev); 2371 virtnet_ack_link_announce(vi); 2372 } 2373 2374 /* Ignore unknown (future) status bits */ 2375 v &= VIRTIO_NET_S_LINK_UP; 2376 2377 if (vi->status == v) 2378 return; 2379 2380 vi->status = v; 2381 2382 if (vi->status & VIRTIO_NET_S_LINK_UP) { 2383 virtnet_update_settings(vi); 2384 netif_carrier_on(vi->dev); 2385 netif_tx_wake_all_queues(vi->dev); 2386 } else { 2387 netif_carrier_off(vi->dev); 2388 netif_tx_stop_all_queues(vi->dev); 2389 } 2390 } 2391 2392 static void virtnet_config_changed(struct virtio_device *vdev) 2393 { 2394 struct virtnet_info *vi = vdev->priv; 2395 2396 schedule_work(&vi->config_work); 2397 } 2398 2399 static void virtnet_free_queues(struct virtnet_info *vi) 2400 { 2401 int i; 2402 2403 for (i = 0; i < vi->max_queue_pairs; i++) { 2404 napi_hash_del(&vi->rq[i].napi); 2405 netif_napi_del(&vi->rq[i].napi); 2406 netif_napi_del(&vi->sq[i].napi); 2407 } 2408 2409 /* We called napi_hash_del() before netif_napi_del(), 2410 * we need to respect an RCU grace period before freeing vi->rq 2411 */ 2412 synchronize_net(); 2413 2414 kfree(vi->rq); 2415 kfree(vi->sq); 2416 kfree(vi->ctrl); 2417 } 2418 2419 static void _free_receive_bufs(struct virtnet_info *vi) 2420 { 2421 struct bpf_prog *old_prog; 2422 int i; 2423 2424 for (i = 0; i < vi->max_queue_pairs; i++) { 2425 while (vi->rq[i].pages) 2426 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 2427 2428 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2429 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 2430 if (old_prog) 2431 bpf_prog_put(old_prog); 2432 } 2433 } 2434 2435 static void free_receive_bufs(struct virtnet_info *vi) 2436 { 2437 rtnl_lock(); 2438 _free_receive_bufs(vi); 2439 rtnl_unlock(); 2440 } 2441 2442 static void free_receive_page_frags(struct virtnet_info *vi) 2443 { 2444 int i; 2445 for (i = 0; i < vi->max_queue_pairs; i++) 2446 if (vi->rq[i].alloc_frag.page) 2447 put_page(vi->rq[i].alloc_frag.page); 2448 } 2449 2450 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) 2451 { 2452 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs)) 2453 return false; 2454 else if (q < vi->curr_queue_pairs) 2455 return true; 2456 else 2457 return false; 2458 } 2459 2460 static void free_unused_bufs(struct virtnet_info *vi) 2461 { 2462 void *buf; 2463 int i; 2464 2465 for (i = 0; i < vi->max_queue_pairs; i++) { 2466 struct virtqueue *vq = vi->sq[i].vq; 2467 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2468 if (!is_xdp_raw_buffer_queue(vi, i)) 2469 dev_kfree_skb(buf); 2470 else 2471 put_page(virt_to_head_page(buf)); 2472 } 2473 } 2474 2475 for (i = 0; i < vi->max_queue_pairs; i++) { 2476 struct virtqueue *vq = vi->rq[i].vq; 2477 2478 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2479 if (vi->mergeable_rx_bufs) { 2480 put_page(virt_to_head_page(buf)); 2481 } else if (vi->big_packets) { 2482 give_pages(&vi->rq[i], buf); 2483 } else { 2484 put_page(virt_to_head_page(buf)); 2485 } 2486 } 2487 } 2488 } 2489 2490 static void virtnet_del_vqs(struct virtnet_info *vi) 2491 { 2492 struct virtio_device *vdev = vi->vdev; 2493 2494 virtnet_clean_affinity(vi, -1); 2495 2496 vdev->config->del_vqs(vdev); 2497 2498 virtnet_free_queues(vi); 2499 } 2500 2501 /* How large should a single buffer be so a queue full of these can fit at 2502 * least one full packet? 2503 * Logic below assumes the mergeable buffer header is used. 2504 */ 2505 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 2506 { 2507 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2508 unsigned int rq_size = virtqueue_get_vring_size(vq); 2509 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 2510 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 2511 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 2512 2513 return max(max(min_buf_len, hdr_len) - hdr_len, 2514 (unsigned int)GOOD_PACKET_LEN); 2515 } 2516 2517 static int virtnet_find_vqs(struct virtnet_info *vi) 2518 { 2519 vq_callback_t **callbacks; 2520 struct virtqueue **vqs; 2521 int ret = -ENOMEM; 2522 int i, total_vqs; 2523 const char **names; 2524 bool *ctx; 2525 2526 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 2527 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 2528 * possible control vq. 2529 */ 2530 total_vqs = vi->max_queue_pairs * 2 + 2531 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 2532 2533 /* Allocate space for find_vqs parameters */ 2534 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL); 2535 if (!vqs) 2536 goto err_vq; 2537 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL); 2538 if (!callbacks) 2539 goto err_callback; 2540 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL); 2541 if (!names) 2542 goto err_names; 2543 if (!vi->big_packets || vi->mergeable_rx_bufs) { 2544 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL); 2545 if (!ctx) 2546 goto err_ctx; 2547 } else { 2548 ctx = NULL; 2549 } 2550 2551 /* Parameters for control virtqueue, if any */ 2552 if (vi->has_cvq) { 2553 callbacks[total_vqs - 1] = NULL; 2554 names[total_vqs - 1] = "control"; 2555 } 2556 2557 /* Allocate/initialize parameters for send/receive virtqueues */ 2558 for (i = 0; i < vi->max_queue_pairs; i++) { 2559 callbacks[rxq2vq(i)] = skb_recv_done; 2560 callbacks[txq2vq(i)] = skb_xmit_done; 2561 sprintf(vi->rq[i].name, "input.%d", i); 2562 sprintf(vi->sq[i].name, "output.%d", i); 2563 names[rxq2vq(i)] = vi->rq[i].name; 2564 names[txq2vq(i)] = vi->sq[i].name; 2565 if (ctx) 2566 ctx[rxq2vq(i)] = true; 2567 } 2568 2569 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks, 2570 names, ctx, NULL); 2571 if (ret) 2572 goto err_find; 2573 2574 if (vi->has_cvq) { 2575 vi->cvq = vqs[total_vqs - 1]; 2576 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 2577 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 2578 } 2579 2580 for (i = 0; i < vi->max_queue_pairs; i++) { 2581 vi->rq[i].vq = vqs[rxq2vq(i)]; 2582 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 2583 vi->sq[i].vq = vqs[txq2vq(i)]; 2584 } 2585 2586 kfree(names); 2587 kfree(callbacks); 2588 kfree(vqs); 2589 kfree(ctx); 2590 2591 return 0; 2592 2593 err_find: 2594 kfree(ctx); 2595 err_ctx: 2596 kfree(names); 2597 err_names: 2598 kfree(callbacks); 2599 err_callback: 2600 kfree(vqs); 2601 err_vq: 2602 return ret; 2603 } 2604 2605 static int virtnet_alloc_queues(struct virtnet_info *vi) 2606 { 2607 int i; 2608 2609 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL); 2610 if (!vi->ctrl) 2611 goto err_ctrl; 2612 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL); 2613 if (!vi->sq) 2614 goto err_sq; 2615 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL); 2616 if (!vi->rq) 2617 goto err_rq; 2618 2619 INIT_DELAYED_WORK(&vi->refill, refill_work); 2620 for (i = 0; i < vi->max_queue_pairs; i++) { 2621 vi->rq[i].pages = NULL; 2622 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll, 2623 napi_weight); 2624 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx, 2625 napi_tx ? napi_weight : 0); 2626 2627 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 2628 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 2629 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 2630 2631 u64_stats_init(&vi->rq[i].stats.syncp); 2632 u64_stats_init(&vi->sq[i].stats.syncp); 2633 } 2634 2635 return 0; 2636 2637 err_rq: 2638 kfree(vi->sq); 2639 err_sq: 2640 kfree(vi->ctrl); 2641 err_ctrl: 2642 return -ENOMEM; 2643 } 2644 2645 static int init_vqs(struct virtnet_info *vi) 2646 { 2647 int ret; 2648 2649 /* Allocate send & receive queues */ 2650 ret = virtnet_alloc_queues(vi); 2651 if (ret) 2652 goto err; 2653 2654 ret = virtnet_find_vqs(vi); 2655 if (ret) 2656 goto err_free; 2657 2658 get_online_cpus(); 2659 virtnet_set_affinity(vi); 2660 put_online_cpus(); 2661 2662 return 0; 2663 2664 err_free: 2665 virtnet_free_queues(vi); 2666 err: 2667 return ret; 2668 } 2669 2670 #ifdef CONFIG_SYSFS 2671 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 2672 char *buf) 2673 { 2674 struct virtnet_info *vi = netdev_priv(queue->dev); 2675 unsigned int queue_index = get_netdev_rx_queue_index(queue); 2676 unsigned int headroom = virtnet_get_headroom(vi); 2677 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 2678 struct ewma_pkt_len *avg; 2679 2680 BUG_ON(queue_index >= vi->max_queue_pairs); 2681 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 2682 return sprintf(buf, "%u\n", 2683 get_mergeable_buf_len(&vi->rq[queue_index], avg, 2684 SKB_DATA_ALIGN(headroom + tailroom))); 2685 } 2686 2687 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 2688 __ATTR_RO(mergeable_rx_buffer_size); 2689 2690 static struct attribute *virtio_net_mrg_rx_attrs[] = { 2691 &mergeable_rx_buffer_size_attribute.attr, 2692 NULL 2693 }; 2694 2695 static const struct attribute_group virtio_net_mrg_rx_group = { 2696 .name = "virtio_net", 2697 .attrs = virtio_net_mrg_rx_attrs 2698 }; 2699 #endif 2700 2701 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 2702 unsigned int fbit, 2703 const char *fname, const char *dname) 2704 { 2705 if (!virtio_has_feature(vdev, fbit)) 2706 return false; 2707 2708 dev_err(&vdev->dev, "device advertises feature %s but not %s", 2709 fname, dname); 2710 2711 return true; 2712 } 2713 2714 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 2715 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 2716 2717 static bool virtnet_validate_features(struct virtio_device *vdev) 2718 { 2719 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 2720 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 2721 "VIRTIO_NET_F_CTRL_VQ") || 2722 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 2723 "VIRTIO_NET_F_CTRL_VQ") || 2724 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 2725 "VIRTIO_NET_F_CTRL_VQ") || 2726 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 2727 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 2728 "VIRTIO_NET_F_CTRL_VQ"))) { 2729 return false; 2730 } 2731 2732 return true; 2733 } 2734 2735 #define MIN_MTU ETH_MIN_MTU 2736 #define MAX_MTU ETH_MAX_MTU 2737 2738 static int virtnet_validate(struct virtio_device *vdev) 2739 { 2740 if (!vdev->config->get) { 2741 dev_err(&vdev->dev, "%s failure: config access disabled\n", 2742 __func__); 2743 return -EINVAL; 2744 } 2745 2746 if (!virtnet_validate_features(vdev)) 2747 return -EINVAL; 2748 2749 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2750 int mtu = virtio_cread16(vdev, 2751 offsetof(struct virtio_net_config, 2752 mtu)); 2753 if (mtu < MIN_MTU) 2754 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 2755 } 2756 2757 return 0; 2758 } 2759 2760 static int virtnet_probe(struct virtio_device *vdev) 2761 { 2762 int i, err = -ENOMEM; 2763 struct net_device *dev; 2764 struct virtnet_info *vi; 2765 u16 max_queue_pairs; 2766 int mtu; 2767 2768 /* Find if host supports multiqueue virtio_net device */ 2769 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ, 2770 struct virtio_net_config, 2771 max_virtqueue_pairs, &max_queue_pairs); 2772 2773 /* We need at least 2 queue's */ 2774 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 2775 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 2776 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2777 max_queue_pairs = 1; 2778 2779 /* Allocate ourselves a network device with room for our info */ 2780 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 2781 if (!dev) 2782 return -ENOMEM; 2783 2784 /* Set up network device as normal. */ 2785 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE; 2786 dev->netdev_ops = &virtnet_netdev; 2787 dev->features = NETIF_F_HIGHDMA; 2788 2789 dev->ethtool_ops = &virtnet_ethtool_ops; 2790 SET_NETDEV_DEV(dev, &vdev->dev); 2791 2792 /* Do we support "hardware" checksums? */ 2793 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 2794 /* This opens up the world of extra features. */ 2795 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2796 if (csum) 2797 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2798 2799 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 2800 dev->hw_features |= NETIF_F_TSO 2801 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 2802 } 2803 /* Individual feature bits: what can host handle? */ 2804 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 2805 dev->hw_features |= NETIF_F_TSO; 2806 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 2807 dev->hw_features |= NETIF_F_TSO6; 2808 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 2809 dev->hw_features |= NETIF_F_TSO_ECN; 2810 2811 dev->features |= NETIF_F_GSO_ROBUST; 2812 2813 if (gso) 2814 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 2815 /* (!csum && gso) case will be fixed by register_netdev() */ 2816 } 2817 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 2818 dev->features |= NETIF_F_RXCSUM; 2819 2820 dev->vlan_features = dev->features; 2821 2822 /* MTU range: 68 - 65535 */ 2823 dev->min_mtu = MIN_MTU; 2824 dev->max_mtu = MAX_MTU; 2825 2826 /* Configuration may specify what MAC to use. Otherwise random. */ 2827 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) 2828 virtio_cread_bytes(vdev, 2829 offsetof(struct virtio_net_config, mac), 2830 dev->dev_addr, dev->addr_len); 2831 else 2832 eth_hw_addr_random(dev); 2833 2834 /* Set up our device-specific information */ 2835 vi = netdev_priv(dev); 2836 vi->dev = dev; 2837 vi->vdev = vdev; 2838 vdev->priv = vi; 2839 2840 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 2841 2842 /* If we can receive ANY GSO packets, we must allocate large ones. */ 2843 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 2844 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) || 2845 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) || 2846 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO)) 2847 vi->big_packets = true; 2848 2849 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) 2850 vi->mergeable_rx_bufs = true; 2851 2852 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 2853 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2854 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2855 else 2856 vi->hdr_len = sizeof(struct virtio_net_hdr); 2857 2858 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 2859 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2860 vi->any_header_sg = true; 2861 2862 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2863 vi->has_cvq = true; 2864 2865 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2866 mtu = virtio_cread16(vdev, 2867 offsetof(struct virtio_net_config, 2868 mtu)); 2869 if (mtu < dev->min_mtu) { 2870 /* Should never trigger: MTU was previously validated 2871 * in virtnet_validate. 2872 */ 2873 dev_err(&vdev->dev, "device MTU appears to have changed " 2874 "it is now %d < %d", mtu, dev->min_mtu); 2875 goto free; 2876 } 2877 2878 dev->mtu = mtu; 2879 dev->max_mtu = mtu; 2880 2881 /* TODO: size buffers correctly in this case. */ 2882 if (dev->mtu > ETH_DATA_LEN) 2883 vi->big_packets = true; 2884 } 2885 2886 if (vi->any_header_sg) 2887 dev->needed_headroom = vi->hdr_len; 2888 2889 /* Enable multiqueue by default */ 2890 if (num_online_cpus() >= max_queue_pairs) 2891 vi->curr_queue_pairs = max_queue_pairs; 2892 else 2893 vi->curr_queue_pairs = num_online_cpus(); 2894 vi->max_queue_pairs = max_queue_pairs; 2895 2896 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 2897 err = init_vqs(vi); 2898 if (err) 2899 goto free; 2900 2901 #ifdef CONFIG_SYSFS 2902 if (vi->mergeable_rx_bufs) 2903 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 2904 #endif 2905 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 2906 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 2907 2908 virtnet_init_settings(dev); 2909 2910 err = register_netdev(dev); 2911 if (err) { 2912 pr_debug("virtio_net: registering device failed\n"); 2913 goto free_vqs; 2914 } 2915 2916 virtio_device_ready(vdev); 2917 2918 err = virtnet_cpu_notif_add(vi); 2919 if (err) { 2920 pr_debug("virtio_net: registering cpu notifier failed\n"); 2921 goto free_unregister_netdev; 2922 } 2923 2924 virtnet_set_queues(vi, vi->curr_queue_pairs); 2925 2926 /* Assume link up if device can't report link status, 2927 otherwise get link status from config. */ 2928 netif_carrier_off(dev); 2929 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 2930 schedule_work(&vi->config_work); 2931 } else { 2932 vi->status = VIRTIO_NET_S_LINK_UP; 2933 virtnet_update_settings(vi); 2934 netif_carrier_on(dev); 2935 } 2936 2937 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 2938 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 2939 set_bit(guest_offloads[i], &vi->guest_offloads); 2940 2941 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 2942 dev->name, max_queue_pairs); 2943 2944 return 0; 2945 2946 free_unregister_netdev: 2947 vi->vdev->config->reset(vdev); 2948 2949 unregister_netdev(dev); 2950 free_vqs: 2951 cancel_delayed_work_sync(&vi->refill); 2952 free_receive_page_frags(vi); 2953 virtnet_del_vqs(vi); 2954 free: 2955 free_netdev(dev); 2956 return err; 2957 } 2958 2959 static void remove_vq_common(struct virtnet_info *vi) 2960 { 2961 vi->vdev->config->reset(vi->vdev); 2962 2963 /* Free unused buffers in both send and recv, if any. */ 2964 free_unused_bufs(vi); 2965 2966 free_receive_bufs(vi); 2967 2968 free_receive_page_frags(vi); 2969 2970 virtnet_del_vqs(vi); 2971 } 2972 2973 static void virtnet_remove(struct virtio_device *vdev) 2974 { 2975 struct virtnet_info *vi = vdev->priv; 2976 2977 virtnet_cpu_notif_remove(vi); 2978 2979 /* Make sure no work handler is accessing the device. */ 2980 flush_work(&vi->config_work); 2981 2982 unregister_netdev(vi->dev); 2983 2984 remove_vq_common(vi); 2985 2986 free_netdev(vi->dev); 2987 } 2988 2989 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 2990 { 2991 struct virtnet_info *vi = vdev->priv; 2992 2993 virtnet_cpu_notif_remove(vi); 2994 virtnet_freeze_down(vdev); 2995 remove_vq_common(vi); 2996 2997 return 0; 2998 } 2999 3000 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 3001 { 3002 struct virtnet_info *vi = vdev->priv; 3003 int err; 3004 3005 err = virtnet_restore_up(vdev); 3006 if (err) 3007 return err; 3008 virtnet_set_queues(vi, vi->curr_queue_pairs); 3009 3010 err = virtnet_cpu_notif_add(vi); 3011 if (err) 3012 return err; 3013 3014 return 0; 3015 } 3016 3017 static struct virtio_device_id id_table[] = { 3018 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 3019 { 0 }, 3020 }; 3021 3022 #define VIRTNET_FEATURES \ 3023 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 3024 VIRTIO_NET_F_MAC, \ 3025 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 3026 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 3027 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 3028 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 3029 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 3030 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 3031 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 3032 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 3033 VIRTIO_NET_F_SPEED_DUPLEX 3034 3035 static unsigned int features[] = { 3036 VIRTNET_FEATURES, 3037 }; 3038 3039 static unsigned int features_legacy[] = { 3040 VIRTNET_FEATURES, 3041 VIRTIO_NET_F_GSO, 3042 VIRTIO_F_ANY_LAYOUT, 3043 }; 3044 3045 static struct virtio_driver virtio_net_driver = { 3046 .feature_table = features, 3047 .feature_table_size = ARRAY_SIZE(features), 3048 .feature_table_legacy = features_legacy, 3049 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 3050 .driver.name = KBUILD_MODNAME, 3051 .driver.owner = THIS_MODULE, 3052 .id_table = id_table, 3053 .validate = virtnet_validate, 3054 .probe = virtnet_probe, 3055 .remove = virtnet_remove, 3056 .config_changed = virtnet_config_changed, 3057 #ifdef CONFIG_PM_SLEEP 3058 .freeze = virtnet_freeze, 3059 .restore = virtnet_restore, 3060 #endif 3061 }; 3062 3063 static __init int virtio_net_driver_init(void) 3064 { 3065 int ret; 3066 3067 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 3068 virtnet_cpu_online, 3069 virtnet_cpu_down_prep); 3070 if (ret < 0) 3071 goto out; 3072 virtionet_online = ret; 3073 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 3074 NULL, virtnet_cpu_dead); 3075 if (ret) 3076 goto err_dead; 3077 3078 ret = register_virtio_driver(&virtio_net_driver); 3079 if (ret) 3080 goto err_virtio; 3081 return 0; 3082 err_virtio: 3083 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 3084 err_dead: 3085 cpuhp_remove_multi_state(virtionet_online); 3086 out: 3087 return ret; 3088 } 3089 module_init(virtio_net_driver_init); 3090 3091 static __exit void virtio_net_driver_exit(void) 3092 { 3093 unregister_virtio_driver(&virtio_net_driver); 3094 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 3095 cpuhp_remove_multi_state(virtionet_online); 3096 } 3097 module_exit(virtio_net_driver_exit); 3098 3099 MODULE_DEVICE_TABLE(virtio, id_table); 3100 MODULE_DESCRIPTION("Virtio network driver"); 3101 MODULE_LICENSE("GPL"); 3102