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 <linux/netdevice.h> 34 #include <linux/pci.h> 35 #include <net/route.h> 36 #include <net/xdp.h> 37 #include <net/net_failover.h> 38 39 static int napi_weight = NAPI_POLL_WEIGHT; 40 module_param(napi_weight, int, 0444); 41 42 static bool csum = true, gso = true, napi_tx; 43 module_param(csum, bool, 0444); 44 module_param(gso, bool, 0444); 45 module_param(napi_tx, bool, 0644); 46 47 /* FIXME: MTU in config. */ 48 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN) 49 #define GOOD_COPY_LEN 128 50 51 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD) 52 53 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */ 54 #define VIRTIO_XDP_HEADROOM 256 55 56 /* RX packet size EWMA. The average packet size is used to determine the packet 57 * buffer size when refilling RX rings. As the entire RX ring may be refilled 58 * at once, the weight is chosen so that the EWMA will be insensitive to short- 59 * term, transient changes in packet size. 60 */ 61 DECLARE_EWMA(pkt_len, 0, 64) 62 63 #define VIRTNET_DRIVER_VERSION "1.0.0" 64 65 static const unsigned long guest_offloads[] = { 66 VIRTIO_NET_F_GUEST_TSO4, 67 VIRTIO_NET_F_GUEST_TSO6, 68 VIRTIO_NET_F_GUEST_ECN, 69 VIRTIO_NET_F_GUEST_UFO 70 }; 71 72 struct virtnet_stat_desc { 73 char desc[ETH_GSTRING_LEN]; 74 size_t offset; 75 }; 76 77 struct virtnet_sq_stats { 78 struct u64_stats_sync syncp; 79 u64 packets; 80 u64 bytes; 81 }; 82 83 struct virtnet_rq_stats { 84 struct u64_stats_sync syncp; 85 u64 packets; 86 u64 bytes; 87 }; 88 89 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m) 90 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m) 91 92 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = { 93 { "packets", VIRTNET_SQ_STAT(packets) }, 94 { "bytes", VIRTNET_SQ_STAT(bytes) }, 95 }; 96 97 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = { 98 { "packets", VIRTNET_RQ_STAT(packets) }, 99 { "bytes", VIRTNET_RQ_STAT(bytes) }, 100 }; 101 102 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc) 103 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc) 104 105 /* Internal representation of a send virtqueue */ 106 struct send_queue { 107 /* Virtqueue associated with this send _queue */ 108 struct virtqueue *vq; 109 110 /* TX: fragments + linear part + virtio header */ 111 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 112 113 /* Name of the send queue: output.$index */ 114 char name[40]; 115 116 struct virtnet_sq_stats stats; 117 118 struct napi_struct napi; 119 }; 120 121 /* Internal representation of a receive virtqueue */ 122 struct receive_queue { 123 /* Virtqueue associated with this receive_queue */ 124 struct virtqueue *vq; 125 126 struct napi_struct napi; 127 128 struct bpf_prog __rcu *xdp_prog; 129 130 struct virtnet_rq_stats stats; 131 132 /* Chain pages by the private ptr. */ 133 struct page *pages; 134 135 /* Average packet length for mergeable receive buffers. */ 136 struct ewma_pkt_len mrg_avg_pkt_len; 137 138 /* Page frag for packet buffer allocation. */ 139 struct page_frag alloc_frag; 140 141 /* RX: fragments + linear part + virtio header */ 142 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 143 144 /* Min single buffer size for mergeable buffers case. */ 145 unsigned int min_buf_len; 146 147 /* Name of this receive queue: input.$index */ 148 char name[40]; 149 150 struct xdp_rxq_info xdp_rxq; 151 }; 152 153 /* Control VQ buffers: protected by the rtnl lock */ 154 struct control_buf { 155 struct virtio_net_ctrl_hdr hdr; 156 virtio_net_ctrl_ack status; 157 struct virtio_net_ctrl_mq mq; 158 u8 promisc; 159 u8 allmulti; 160 __virtio16 vid; 161 __virtio64 offloads; 162 }; 163 164 struct virtnet_info { 165 struct virtio_device *vdev; 166 struct virtqueue *cvq; 167 struct net_device *dev; 168 struct send_queue *sq; 169 struct receive_queue *rq; 170 unsigned int status; 171 172 /* Max # of queue pairs supported by the device */ 173 u16 max_queue_pairs; 174 175 /* # of queue pairs currently used by the driver */ 176 u16 curr_queue_pairs; 177 178 /* # of XDP queue pairs currently used by the driver */ 179 u16 xdp_queue_pairs; 180 181 /* I like... big packets and I cannot lie! */ 182 bool big_packets; 183 184 /* Host will merge rx buffers for big packets (shake it! shake it!) */ 185 bool mergeable_rx_bufs; 186 187 /* Has control virtqueue */ 188 bool has_cvq; 189 190 /* Host can handle any s/g split between our header and packet data */ 191 bool any_header_sg; 192 193 /* Packet virtio header size */ 194 u8 hdr_len; 195 196 /* Work struct for refilling if we run low on memory. */ 197 struct delayed_work refill; 198 199 /* Work struct for config space updates */ 200 struct work_struct config_work; 201 202 /* Does the affinity hint is set for virtqueues? */ 203 bool affinity_hint_set; 204 205 /* CPU hotplug instances for online & dead */ 206 struct hlist_node node; 207 struct hlist_node node_dead; 208 209 struct control_buf *ctrl; 210 211 /* Ethtool settings */ 212 u8 duplex; 213 u32 speed; 214 215 unsigned long guest_offloads; 216 217 /* failover when STANDBY feature enabled */ 218 struct failover *failover; 219 }; 220 221 struct padded_vnet_hdr { 222 struct virtio_net_hdr_mrg_rxbuf hdr; 223 /* 224 * hdr is in a separate sg buffer, and data sg buffer shares same page 225 * with this header sg. This padding makes next sg 16 byte aligned 226 * after the header. 227 */ 228 char padding[4]; 229 }; 230 231 /* Converting between virtqueue no. and kernel tx/rx queue no. 232 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq 233 */ 234 static int vq2txq(struct virtqueue *vq) 235 { 236 return (vq->index - 1) / 2; 237 } 238 239 static int txq2vq(int txq) 240 { 241 return txq * 2 + 1; 242 } 243 244 static int vq2rxq(struct virtqueue *vq) 245 { 246 return vq->index / 2; 247 } 248 249 static int rxq2vq(int rxq) 250 { 251 return rxq * 2; 252 } 253 254 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb) 255 { 256 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb; 257 } 258 259 /* 260 * private is used to chain pages for big packets, put the whole 261 * most recent used list in the beginning for reuse 262 */ 263 static void give_pages(struct receive_queue *rq, struct page *page) 264 { 265 struct page *end; 266 267 /* Find end of list, sew whole thing into vi->rq.pages. */ 268 for (end = page; end->private; end = (struct page *)end->private); 269 end->private = (unsigned long)rq->pages; 270 rq->pages = page; 271 } 272 273 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask) 274 { 275 struct page *p = rq->pages; 276 277 if (p) { 278 rq->pages = (struct page *)p->private; 279 /* clear private here, it is used to chain pages */ 280 p->private = 0; 281 } else 282 p = alloc_page(gfp_mask); 283 return p; 284 } 285 286 static void virtqueue_napi_schedule(struct napi_struct *napi, 287 struct virtqueue *vq) 288 { 289 if (napi_schedule_prep(napi)) { 290 virtqueue_disable_cb(vq); 291 __napi_schedule(napi); 292 } 293 } 294 295 static void virtqueue_napi_complete(struct napi_struct *napi, 296 struct virtqueue *vq, int processed) 297 { 298 int opaque; 299 300 opaque = virtqueue_enable_cb_prepare(vq); 301 if (napi_complete_done(napi, processed)) { 302 if (unlikely(virtqueue_poll(vq, opaque))) 303 virtqueue_napi_schedule(napi, vq); 304 } else { 305 virtqueue_disable_cb(vq); 306 } 307 } 308 309 static void skb_xmit_done(struct virtqueue *vq) 310 { 311 struct virtnet_info *vi = vq->vdev->priv; 312 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi; 313 314 /* Suppress further interrupts. */ 315 virtqueue_disable_cb(vq); 316 317 if (napi->weight) 318 virtqueue_napi_schedule(napi, vq); 319 else 320 /* We were probably waiting for more output buffers. */ 321 netif_wake_subqueue(vi->dev, vq2txq(vq)); 322 } 323 324 #define MRG_CTX_HEADER_SHIFT 22 325 static void *mergeable_len_to_ctx(unsigned int truesize, 326 unsigned int headroom) 327 { 328 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize); 329 } 330 331 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx) 332 { 333 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT; 334 } 335 336 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx) 337 { 338 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1); 339 } 340 341 /* Called from bottom half context */ 342 static struct sk_buff *page_to_skb(struct virtnet_info *vi, 343 struct receive_queue *rq, 344 struct page *page, unsigned int offset, 345 unsigned int len, unsigned int truesize) 346 { 347 struct sk_buff *skb; 348 struct virtio_net_hdr_mrg_rxbuf *hdr; 349 unsigned int copy, hdr_len, hdr_padded_len; 350 char *p; 351 352 p = page_address(page) + offset; 353 354 /* copy small packet so we can reuse these pages for small data */ 355 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN); 356 if (unlikely(!skb)) 357 return NULL; 358 359 hdr = skb_vnet_hdr(skb); 360 361 hdr_len = vi->hdr_len; 362 if (vi->mergeable_rx_bufs) 363 hdr_padded_len = sizeof(*hdr); 364 else 365 hdr_padded_len = sizeof(struct padded_vnet_hdr); 366 367 memcpy(hdr, p, hdr_len); 368 369 len -= hdr_len; 370 offset += hdr_padded_len; 371 p += hdr_padded_len; 372 373 copy = len; 374 if (copy > skb_tailroom(skb)) 375 copy = skb_tailroom(skb); 376 skb_put_data(skb, p, copy); 377 378 len -= copy; 379 offset += copy; 380 381 if (vi->mergeable_rx_bufs) { 382 if (len) 383 skb_add_rx_frag(skb, 0, page, offset, len, truesize); 384 else 385 put_page(page); 386 return skb; 387 } 388 389 /* 390 * Verify that we can indeed put this data into a skb. 391 * This is here to handle cases when the device erroneously 392 * tries to receive more than is possible. This is usually 393 * the case of a broken device. 394 */ 395 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) { 396 net_dbg_ratelimited("%s: too much data\n", skb->dev->name); 397 dev_kfree_skb(skb); 398 return NULL; 399 } 400 BUG_ON(offset >= PAGE_SIZE); 401 while (len) { 402 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len); 403 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, 404 frag_size, truesize); 405 len -= frag_size; 406 page = (struct page *)page->private; 407 offset = 0; 408 } 409 410 if (page) 411 give_pages(rq, page); 412 413 return skb; 414 } 415 416 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 417 struct send_queue *sq, 418 struct xdp_frame *xdpf) 419 { 420 struct virtio_net_hdr_mrg_rxbuf *hdr; 421 int err; 422 423 /* virtqueue want to use data area in-front of packet */ 424 if (unlikely(xdpf->metasize > 0)) 425 return -EOPNOTSUPP; 426 427 if (unlikely(xdpf->headroom < vi->hdr_len)) 428 return -EOVERFLOW; 429 430 /* Make room for virtqueue hdr (also change xdpf->headroom?) */ 431 xdpf->data -= vi->hdr_len; 432 /* Zero header and leave csum up to XDP layers */ 433 hdr = xdpf->data; 434 memset(hdr, 0, vi->hdr_len); 435 xdpf->len += vi->hdr_len; 436 437 sg_init_one(sq->sg, xdpf->data, xdpf->len); 438 439 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC); 440 if (unlikely(err)) 441 return -ENOSPC; /* Caller handle free/refcnt */ 442 443 return 0; 444 } 445 446 static int __virtnet_xdp_tx_xmit(struct virtnet_info *vi, 447 struct xdp_frame *xdpf) 448 { 449 struct xdp_frame *xdpf_sent; 450 struct send_queue *sq; 451 unsigned int len; 452 unsigned int qp; 453 454 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 455 sq = &vi->sq[qp]; 456 457 /* Free up any pending old buffers before queueing new ones. */ 458 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) 459 xdp_return_frame(xdpf_sent); 460 461 return __virtnet_xdp_xmit_one(vi, sq, xdpf); 462 } 463 464 static int virtnet_xdp_xmit(struct net_device *dev, 465 int n, struct xdp_frame **frames, u32 flags) 466 { 467 struct virtnet_info *vi = netdev_priv(dev); 468 struct receive_queue *rq = vi->rq; 469 struct xdp_frame *xdpf_sent; 470 struct bpf_prog *xdp_prog; 471 struct send_queue *sq; 472 unsigned int len; 473 unsigned int qp; 474 int drops = 0; 475 int err; 476 int i; 477 478 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) 479 return -EINVAL; 480 481 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 482 sq = &vi->sq[qp]; 483 484 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this 485 * indicate XDP resources have been successfully allocated. 486 */ 487 xdp_prog = rcu_dereference(rq->xdp_prog); 488 if (!xdp_prog) 489 return -ENXIO; 490 491 /* Free up any pending old buffers before queueing new ones. */ 492 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) 493 xdp_return_frame(xdpf_sent); 494 495 for (i = 0; i < n; i++) { 496 struct xdp_frame *xdpf = frames[i]; 497 498 err = __virtnet_xdp_xmit_one(vi, sq, xdpf); 499 if (err) { 500 xdp_return_frame_rx_napi(xdpf); 501 drops++; 502 } 503 } 504 505 if (flags & XDP_XMIT_FLUSH) 506 virtqueue_kick(sq->vq); 507 508 return n - drops; 509 } 510 511 static unsigned int virtnet_get_headroom(struct virtnet_info *vi) 512 { 513 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0; 514 } 515 516 /* We copy the packet for XDP in the following cases: 517 * 518 * 1) Packet is scattered across multiple rx buffers. 519 * 2) Headroom space is insufficient. 520 * 521 * This is inefficient but it's a temporary condition that 522 * we hit right after XDP is enabled and until queue is refilled 523 * with large buffers with sufficient headroom - so it should affect 524 * at most queue size packets. 525 * Afterwards, the conditions to enable 526 * XDP should preclude the underlying device from sending packets 527 * across multiple buffers (num_buf > 1), and we make sure buffers 528 * have enough headroom. 529 */ 530 static struct page *xdp_linearize_page(struct receive_queue *rq, 531 u16 *num_buf, 532 struct page *p, 533 int offset, 534 int page_off, 535 unsigned int *len) 536 { 537 struct page *page = alloc_page(GFP_ATOMIC); 538 539 if (!page) 540 return NULL; 541 542 memcpy(page_address(page) + page_off, page_address(p) + offset, *len); 543 page_off += *len; 544 545 while (--*num_buf) { 546 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 547 unsigned int buflen; 548 void *buf; 549 int off; 550 551 buf = virtqueue_get_buf(rq->vq, &buflen); 552 if (unlikely(!buf)) 553 goto err_buf; 554 555 p = virt_to_head_page(buf); 556 off = buf - page_address(p); 557 558 /* guard against a misconfigured or uncooperative backend that 559 * is sending packet larger than the MTU. 560 */ 561 if ((page_off + buflen + tailroom) > PAGE_SIZE) { 562 put_page(p); 563 goto err_buf; 564 } 565 566 memcpy(page_address(page) + page_off, 567 page_address(p) + off, buflen); 568 page_off += buflen; 569 put_page(p); 570 } 571 572 /* Headroom does not contribute to packet length */ 573 *len = page_off - VIRTIO_XDP_HEADROOM; 574 return page; 575 err_buf: 576 __free_pages(page, 0); 577 return NULL; 578 } 579 580 static struct sk_buff *receive_small(struct net_device *dev, 581 struct virtnet_info *vi, 582 struct receive_queue *rq, 583 void *buf, void *ctx, 584 unsigned int len, 585 bool *xdp_xmit) 586 { 587 struct sk_buff *skb; 588 struct bpf_prog *xdp_prog; 589 unsigned int xdp_headroom = (unsigned long)ctx; 590 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom; 591 unsigned int headroom = vi->hdr_len + header_offset; 592 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 593 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 594 struct page *page = virt_to_head_page(buf); 595 unsigned int delta = 0; 596 struct page *xdp_page; 597 int err; 598 599 len -= vi->hdr_len; 600 601 rcu_read_lock(); 602 xdp_prog = rcu_dereference(rq->xdp_prog); 603 if (xdp_prog) { 604 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset; 605 struct xdp_frame *xdpf; 606 struct xdp_buff xdp; 607 void *orig_data; 608 u32 act; 609 610 if (unlikely(hdr->hdr.gso_type)) 611 goto err_xdp; 612 613 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) { 614 int offset = buf - page_address(page) + header_offset; 615 unsigned int tlen = len + vi->hdr_len; 616 u16 num_buf = 1; 617 618 xdp_headroom = virtnet_get_headroom(vi); 619 header_offset = VIRTNET_RX_PAD + xdp_headroom; 620 headroom = vi->hdr_len + header_offset; 621 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 622 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 623 xdp_page = xdp_linearize_page(rq, &num_buf, page, 624 offset, header_offset, 625 &tlen); 626 if (!xdp_page) 627 goto err_xdp; 628 629 buf = page_address(xdp_page); 630 put_page(page); 631 page = xdp_page; 632 } 633 634 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len; 635 xdp.data = xdp.data_hard_start + xdp_headroom; 636 xdp_set_data_meta_invalid(&xdp); 637 xdp.data_end = xdp.data + len; 638 xdp.rxq = &rq->xdp_rxq; 639 orig_data = xdp.data; 640 act = bpf_prog_run_xdp(xdp_prog, &xdp); 641 642 switch (act) { 643 case XDP_PASS: 644 /* Recalculate length in case bpf program changed it */ 645 delta = orig_data - xdp.data; 646 len = xdp.data_end - xdp.data; 647 break; 648 case XDP_TX: 649 xdpf = convert_to_xdp_frame(&xdp); 650 if (unlikely(!xdpf)) 651 goto err_xdp; 652 err = __virtnet_xdp_tx_xmit(vi, xdpf); 653 if (unlikely(err)) { 654 trace_xdp_exception(vi->dev, xdp_prog, act); 655 goto err_xdp; 656 } 657 *xdp_xmit = true; 658 rcu_read_unlock(); 659 goto xdp_xmit; 660 case XDP_REDIRECT: 661 err = xdp_do_redirect(dev, &xdp, xdp_prog); 662 if (err) 663 goto err_xdp; 664 *xdp_xmit = true; 665 rcu_read_unlock(); 666 goto xdp_xmit; 667 default: 668 bpf_warn_invalid_xdp_action(act); 669 case XDP_ABORTED: 670 trace_xdp_exception(vi->dev, xdp_prog, act); 671 case XDP_DROP: 672 goto err_xdp; 673 } 674 } 675 rcu_read_unlock(); 676 677 skb = build_skb(buf, buflen); 678 if (!skb) { 679 put_page(page); 680 goto err; 681 } 682 skb_reserve(skb, headroom - delta); 683 skb_put(skb, len); 684 if (!delta) { 685 buf += header_offset; 686 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len); 687 } /* keep zeroed vnet hdr since packet was changed by bpf */ 688 689 err: 690 return skb; 691 692 err_xdp: 693 rcu_read_unlock(); 694 dev->stats.rx_dropped++; 695 put_page(page); 696 xdp_xmit: 697 return NULL; 698 } 699 700 static struct sk_buff *receive_big(struct net_device *dev, 701 struct virtnet_info *vi, 702 struct receive_queue *rq, 703 void *buf, 704 unsigned int len) 705 { 706 struct page *page = buf; 707 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE); 708 709 if (unlikely(!skb)) 710 goto err; 711 712 return skb; 713 714 err: 715 dev->stats.rx_dropped++; 716 give_pages(rq, page); 717 return NULL; 718 } 719 720 static struct sk_buff *receive_mergeable(struct net_device *dev, 721 struct virtnet_info *vi, 722 struct receive_queue *rq, 723 void *buf, 724 void *ctx, 725 unsigned int len, 726 bool *xdp_xmit) 727 { 728 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 729 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 730 struct page *page = virt_to_head_page(buf); 731 int offset = buf - page_address(page); 732 struct sk_buff *head_skb, *curr_skb; 733 struct bpf_prog *xdp_prog; 734 unsigned int truesize; 735 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 736 int err; 737 738 head_skb = NULL; 739 740 rcu_read_lock(); 741 xdp_prog = rcu_dereference(rq->xdp_prog); 742 if (xdp_prog) { 743 struct xdp_frame *xdpf; 744 struct page *xdp_page; 745 struct xdp_buff xdp; 746 void *data; 747 u32 act; 748 749 /* Transient failure which in theory could occur if 750 * in-flight packets from before XDP was enabled reach 751 * the receive path after XDP is loaded. 752 */ 753 if (unlikely(hdr->hdr.gso_type)) 754 goto err_xdp; 755 756 /* This happens when rx buffer size is underestimated 757 * or headroom is not enough because of the buffer 758 * was refilled before XDP is set. This should only 759 * happen for the first several packets, so we don't 760 * care much about its performance. 761 */ 762 if (unlikely(num_buf > 1 || 763 headroom < virtnet_get_headroom(vi))) { 764 /* linearize data for XDP */ 765 xdp_page = xdp_linearize_page(rq, &num_buf, 766 page, offset, 767 VIRTIO_XDP_HEADROOM, 768 &len); 769 if (!xdp_page) 770 goto err_xdp; 771 offset = VIRTIO_XDP_HEADROOM; 772 } else { 773 xdp_page = page; 774 } 775 776 /* Allow consuming headroom but reserve enough space to push 777 * the descriptor on if we get an XDP_TX return code. 778 */ 779 data = page_address(xdp_page) + offset; 780 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len; 781 xdp.data = data + vi->hdr_len; 782 xdp_set_data_meta_invalid(&xdp); 783 xdp.data_end = xdp.data + (len - vi->hdr_len); 784 xdp.rxq = &rq->xdp_rxq; 785 786 act = bpf_prog_run_xdp(xdp_prog, &xdp); 787 788 switch (act) { 789 case XDP_PASS: 790 /* recalculate offset to account for any header 791 * adjustments. Note other cases do not build an 792 * skb and avoid using offset 793 */ 794 offset = xdp.data - 795 page_address(xdp_page) - vi->hdr_len; 796 797 /* recalculate len if xdp.data or xdp.data_end were 798 * adjusted 799 */ 800 len = xdp.data_end - xdp.data + vi->hdr_len; 801 /* We can only create skb based on xdp_page. */ 802 if (unlikely(xdp_page != page)) { 803 rcu_read_unlock(); 804 put_page(page); 805 head_skb = page_to_skb(vi, rq, xdp_page, 806 offset, len, PAGE_SIZE); 807 return head_skb; 808 } 809 break; 810 case XDP_TX: 811 xdpf = convert_to_xdp_frame(&xdp); 812 if (unlikely(!xdpf)) 813 goto err_xdp; 814 err = __virtnet_xdp_tx_xmit(vi, xdpf); 815 if (unlikely(err)) { 816 trace_xdp_exception(vi->dev, xdp_prog, act); 817 if (unlikely(xdp_page != page)) 818 put_page(xdp_page); 819 goto err_xdp; 820 } 821 *xdp_xmit = true; 822 if (unlikely(xdp_page != page)) 823 put_page(page); 824 rcu_read_unlock(); 825 goto xdp_xmit; 826 case XDP_REDIRECT: 827 err = xdp_do_redirect(dev, &xdp, xdp_prog); 828 if (err) { 829 if (unlikely(xdp_page != page)) 830 put_page(xdp_page); 831 goto err_xdp; 832 } 833 *xdp_xmit = true; 834 if (unlikely(xdp_page != page)) 835 put_page(page); 836 rcu_read_unlock(); 837 goto xdp_xmit; 838 default: 839 bpf_warn_invalid_xdp_action(act); 840 case XDP_ABORTED: 841 trace_xdp_exception(vi->dev, xdp_prog, act); 842 case XDP_DROP: 843 if (unlikely(xdp_page != page)) 844 __free_pages(xdp_page, 0); 845 goto err_xdp; 846 } 847 } 848 rcu_read_unlock(); 849 850 truesize = mergeable_ctx_to_truesize(ctx); 851 if (unlikely(len > truesize)) { 852 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 853 dev->name, len, (unsigned long)ctx); 854 dev->stats.rx_length_errors++; 855 goto err_skb; 856 } 857 858 head_skb = page_to_skb(vi, rq, page, offset, len, truesize); 859 curr_skb = head_skb; 860 861 if (unlikely(!curr_skb)) 862 goto err_skb; 863 while (--num_buf) { 864 int num_skb_frags; 865 866 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 867 if (unlikely(!buf)) { 868 pr_debug("%s: rx error: %d buffers out of %d missing\n", 869 dev->name, num_buf, 870 virtio16_to_cpu(vi->vdev, 871 hdr->num_buffers)); 872 dev->stats.rx_length_errors++; 873 goto err_buf; 874 } 875 876 page = virt_to_head_page(buf); 877 878 truesize = mergeable_ctx_to_truesize(ctx); 879 if (unlikely(len > truesize)) { 880 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 881 dev->name, len, (unsigned long)ctx); 882 dev->stats.rx_length_errors++; 883 goto err_skb; 884 } 885 886 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 887 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 888 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 889 890 if (unlikely(!nskb)) 891 goto err_skb; 892 if (curr_skb == head_skb) 893 skb_shinfo(curr_skb)->frag_list = nskb; 894 else 895 curr_skb->next = nskb; 896 curr_skb = nskb; 897 head_skb->truesize += nskb->truesize; 898 num_skb_frags = 0; 899 } 900 if (curr_skb != head_skb) { 901 head_skb->data_len += len; 902 head_skb->len += len; 903 head_skb->truesize += truesize; 904 } 905 offset = buf - page_address(page); 906 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 907 put_page(page); 908 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 909 len, truesize); 910 } else { 911 skb_add_rx_frag(curr_skb, num_skb_frags, page, 912 offset, len, truesize); 913 } 914 } 915 916 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 917 return head_skb; 918 919 err_xdp: 920 rcu_read_unlock(); 921 err_skb: 922 put_page(page); 923 while (num_buf-- > 1) { 924 buf = virtqueue_get_buf(rq->vq, &len); 925 if (unlikely(!buf)) { 926 pr_debug("%s: rx error: %d buffers missing\n", 927 dev->name, num_buf); 928 dev->stats.rx_length_errors++; 929 break; 930 } 931 page = virt_to_head_page(buf); 932 put_page(page); 933 } 934 err_buf: 935 dev->stats.rx_dropped++; 936 dev_kfree_skb(head_skb); 937 xdp_xmit: 938 return NULL; 939 } 940 941 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 942 void *buf, unsigned int len, void **ctx, bool *xdp_xmit) 943 { 944 struct net_device *dev = vi->dev; 945 struct sk_buff *skb; 946 struct virtio_net_hdr_mrg_rxbuf *hdr; 947 int ret; 948 949 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 950 pr_debug("%s: short packet %i\n", dev->name, len); 951 dev->stats.rx_length_errors++; 952 if (vi->mergeable_rx_bufs) { 953 put_page(virt_to_head_page(buf)); 954 } else if (vi->big_packets) { 955 give_pages(rq, buf); 956 } else { 957 put_page(virt_to_head_page(buf)); 958 } 959 return 0; 960 } 961 962 if (vi->mergeable_rx_bufs) 963 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit); 964 else if (vi->big_packets) 965 skb = receive_big(dev, vi, rq, buf, len); 966 else 967 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit); 968 969 if (unlikely(!skb)) 970 return 0; 971 972 hdr = skb_vnet_hdr(skb); 973 974 ret = skb->len; 975 976 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 977 skb->ip_summed = CHECKSUM_UNNECESSARY; 978 979 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 980 virtio_is_little_endian(vi->vdev))) { 981 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 982 dev->name, hdr->hdr.gso_type, 983 hdr->hdr.gso_size); 984 goto frame_err; 985 } 986 987 skb->protocol = eth_type_trans(skb, dev); 988 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 989 ntohs(skb->protocol), skb->len, skb->pkt_type); 990 991 napi_gro_receive(&rq->napi, skb); 992 return ret; 993 994 frame_err: 995 dev->stats.rx_frame_errors++; 996 dev_kfree_skb(skb); 997 return 0; 998 } 999 1000 /* Unlike mergeable buffers, all buffers are allocated to the 1001 * same size, except for the headroom. For this reason we do 1002 * not need to use mergeable_len_to_ctx here - it is enough 1003 * to store the headroom as the context ignoring the truesize. 1004 */ 1005 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 1006 gfp_t gfp) 1007 { 1008 struct page_frag *alloc_frag = &rq->alloc_frag; 1009 char *buf; 1010 unsigned int xdp_headroom = virtnet_get_headroom(vi); 1011 void *ctx = (void *)(unsigned long)xdp_headroom; 1012 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 1013 int err; 1014 1015 len = SKB_DATA_ALIGN(len) + 1016 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1017 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 1018 return -ENOMEM; 1019 1020 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1021 get_page(alloc_frag->page); 1022 alloc_frag->offset += len; 1023 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 1024 vi->hdr_len + GOOD_PACKET_LEN); 1025 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1026 if (err < 0) 1027 put_page(virt_to_head_page(buf)); 1028 return err; 1029 } 1030 1031 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 1032 gfp_t gfp) 1033 { 1034 struct page *first, *list = NULL; 1035 char *p; 1036 int i, err, offset; 1037 1038 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2); 1039 1040 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */ 1041 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) { 1042 first = get_a_page(rq, gfp); 1043 if (!first) { 1044 if (list) 1045 give_pages(rq, list); 1046 return -ENOMEM; 1047 } 1048 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 1049 1050 /* chain new page in list head to match sg */ 1051 first->private = (unsigned long)list; 1052 list = first; 1053 } 1054 1055 first = get_a_page(rq, gfp); 1056 if (!first) { 1057 give_pages(rq, list); 1058 return -ENOMEM; 1059 } 1060 p = page_address(first); 1061 1062 /* rq->sg[0], rq->sg[1] share the same page */ 1063 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 1064 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 1065 1066 /* rq->sg[1] for data packet, from offset */ 1067 offset = sizeof(struct padded_vnet_hdr); 1068 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 1069 1070 /* chain first in list head */ 1071 first->private = (unsigned long)list; 1072 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2, 1073 first, gfp); 1074 if (err < 0) 1075 give_pages(rq, first); 1076 1077 return err; 1078 } 1079 1080 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 1081 struct ewma_pkt_len *avg_pkt_len, 1082 unsigned int room) 1083 { 1084 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 1085 unsigned int len; 1086 1087 if (room) 1088 return PAGE_SIZE - room; 1089 1090 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1091 rq->min_buf_len, PAGE_SIZE - hdr_len); 1092 1093 return ALIGN(len, L1_CACHE_BYTES); 1094 } 1095 1096 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1097 struct receive_queue *rq, gfp_t gfp) 1098 { 1099 struct page_frag *alloc_frag = &rq->alloc_frag; 1100 unsigned int headroom = virtnet_get_headroom(vi); 1101 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1102 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1103 char *buf; 1104 void *ctx; 1105 int err; 1106 unsigned int len, hole; 1107 1108 /* Extra tailroom is needed to satisfy XDP's assumption. This 1109 * means rx frags coalescing won't work, but consider we've 1110 * disabled GSO for XDP, it won't be a big issue. 1111 */ 1112 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room); 1113 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp))) 1114 return -ENOMEM; 1115 1116 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1117 buf += headroom; /* advance address leaving hole at front of pkt */ 1118 get_page(alloc_frag->page); 1119 alloc_frag->offset += len + room; 1120 hole = alloc_frag->size - alloc_frag->offset; 1121 if (hole < len + room) { 1122 /* To avoid internal fragmentation, if there is very likely not 1123 * enough space for another buffer, add the remaining space to 1124 * the current buffer. 1125 */ 1126 len += hole; 1127 alloc_frag->offset += hole; 1128 } 1129 1130 sg_init_one(rq->sg, buf, len); 1131 ctx = mergeable_len_to_ctx(len, headroom); 1132 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1133 if (err < 0) 1134 put_page(virt_to_head_page(buf)); 1135 1136 return err; 1137 } 1138 1139 /* 1140 * Returns false if we couldn't fill entirely (OOM). 1141 * 1142 * Normally run in the receive path, but can also be run from ndo_open 1143 * before we're receiving packets, or from refill_work which is 1144 * careful to disable receiving (using napi_disable). 1145 */ 1146 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1147 gfp_t gfp) 1148 { 1149 int err; 1150 bool oom; 1151 1152 do { 1153 if (vi->mergeable_rx_bufs) 1154 err = add_recvbuf_mergeable(vi, rq, gfp); 1155 else if (vi->big_packets) 1156 err = add_recvbuf_big(vi, rq, gfp); 1157 else 1158 err = add_recvbuf_small(vi, rq, gfp); 1159 1160 oom = err == -ENOMEM; 1161 if (err) 1162 break; 1163 } while (rq->vq->num_free); 1164 virtqueue_kick(rq->vq); 1165 return !oom; 1166 } 1167 1168 static void skb_recv_done(struct virtqueue *rvq) 1169 { 1170 struct virtnet_info *vi = rvq->vdev->priv; 1171 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1172 1173 virtqueue_napi_schedule(&rq->napi, rvq); 1174 } 1175 1176 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1177 { 1178 napi_enable(napi); 1179 1180 /* If all buffers were filled by other side before we napi_enabled, we 1181 * won't get another interrupt, so process any outstanding packets now. 1182 * Call local_bh_enable after to trigger softIRQ processing. 1183 */ 1184 local_bh_disable(); 1185 virtqueue_napi_schedule(napi, vq); 1186 local_bh_enable(); 1187 } 1188 1189 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1190 struct virtqueue *vq, 1191 struct napi_struct *napi) 1192 { 1193 if (!napi->weight) 1194 return; 1195 1196 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1197 * enable the feature if this is likely affine with the transmit path. 1198 */ 1199 if (!vi->affinity_hint_set) { 1200 napi->weight = 0; 1201 return; 1202 } 1203 1204 return virtnet_napi_enable(vq, napi); 1205 } 1206 1207 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1208 { 1209 if (napi->weight) 1210 napi_disable(napi); 1211 } 1212 1213 static void refill_work(struct work_struct *work) 1214 { 1215 struct virtnet_info *vi = 1216 container_of(work, struct virtnet_info, refill.work); 1217 bool still_empty; 1218 int i; 1219 1220 for (i = 0; i < vi->curr_queue_pairs; i++) { 1221 struct receive_queue *rq = &vi->rq[i]; 1222 1223 napi_disable(&rq->napi); 1224 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1225 virtnet_napi_enable(rq->vq, &rq->napi); 1226 1227 /* In theory, this can happen: if we don't get any buffers in 1228 * we will *never* try to fill again. 1229 */ 1230 if (still_empty) 1231 schedule_delayed_work(&vi->refill, HZ/2); 1232 } 1233 } 1234 1235 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit) 1236 { 1237 struct virtnet_info *vi = rq->vq->vdev->priv; 1238 unsigned int len, received = 0, bytes = 0; 1239 void *buf; 1240 1241 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1242 void *ctx; 1243 1244 while (received < budget && 1245 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1246 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit); 1247 received++; 1248 } 1249 } else { 1250 while (received < budget && 1251 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1252 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit); 1253 received++; 1254 } 1255 } 1256 1257 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) { 1258 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) 1259 schedule_delayed_work(&vi->refill, 0); 1260 } 1261 1262 u64_stats_update_begin(&rq->stats.syncp); 1263 rq->stats.bytes += bytes; 1264 rq->stats.packets += received; 1265 u64_stats_update_end(&rq->stats.syncp); 1266 1267 return received; 1268 } 1269 1270 static void free_old_xmit_skbs(struct send_queue *sq) 1271 { 1272 struct sk_buff *skb; 1273 unsigned int len; 1274 unsigned int packets = 0; 1275 unsigned int bytes = 0; 1276 1277 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) { 1278 pr_debug("Sent skb %p\n", skb); 1279 1280 bytes += skb->len; 1281 packets++; 1282 1283 dev_consume_skb_any(skb); 1284 } 1285 1286 /* Avoid overhead when no packets have been processed 1287 * happens when called speculatively from start_xmit. 1288 */ 1289 if (!packets) 1290 return; 1291 1292 u64_stats_update_begin(&sq->stats.syncp); 1293 sq->stats.bytes += bytes; 1294 sq->stats.packets += packets; 1295 u64_stats_update_end(&sq->stats.syncp); 1296 } 1297 1298 static void virtnet_poll_cleantx(struct receive_queue *rq) 1299 { 1300 struct virtnet_info *vi = rq->vq->vdev->priv; 1301 unsigned int index = vq2rxq(rq->vq); 1302 struct send_queue *sq = &vi->sq[index]; 1303 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1304 1305 if (!sq->napi.weight) 1306 return; 1307 1308 if (__netif_tx_trylock(txq)) { 1309 free_old_xmit_skbs(sq); 1310 __netif_tx_unlock(txq); 1311 } 1312 1313 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1314 netif_tx_wake_queue(txq); 1315 } 1316 1317 static int virtnet_poll(struct napi_struct *napi, int budget) 1318 { 1319 struct receive_queue *rq = 1320 container_of(napi, struct receive_queue, napi); 1321 struct virtnet_info *vi = rq->vq->vdev->priv; 1322 struct send_queue *sq; 1323 unsigned int received, qp; 1324 bool xdp_xmit = false; 1325 1326 virtnet_poll_cleantx(rq); 1327 1328 received = virtnet_receive(rq, budget, &xdp_xmit); 1329 1330 /* Out of packets? */ 1331 if (received < budget) 1332 virtqueue_napi_complete(napi, rq->vq, received); 1333 1334 if (xdp_xmit) { 1335 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + 1336 smp_processor_id(); 1337 sq = &vi->sq[qp]; 1338 virtqueue_kick(sq->vq); 1339 xdp_do_flush_map(); 1340 } 1341 1342 return received; 1343 } 1344 1345 static int virtnet_open(struct net_device *dev) 1346 { 1347 struct virtnet_info *vi = netdev_priv(dev); 1348 int i, err; 1349 1350 for (i = 0; i < vi->max_queue_pairs; i++) { 1351 if (i < vi->curr_queue_pairs) 1352 /* Make sure we have some buffers: if oom use wq. */ 1353 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1354 schedule_delayed_work(&vi->refill, 0); 1355 1356 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i); 1357 if (err < 0) 1358 return err; 1359 1360 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq, 1361 MEM_TYPE_PAGE_SHARED, NULL); 1362 if (err < 0) { 1363 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1364 return err; 1365 } 1366 1367 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 1368 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 1369 } 1370 1371 return 0; 1372 } 1373 1374 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 1375 { 1376 struct send_queue *sq = container_of(napi, struct send_queue, napi); 1377 struct virtnet_info *vi = sq->vq->vdev->priv; 1378 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq)); 1379 1380 __netif_tx_lock(txq, raw_smp_processor_id()); 1381 free_old_xmit_skbs(sq); 1382 __netif_tx_unlock(txq); 1383 1384 virtqueue_napi_complete(napi, sq->vq, 0); 1385 1386 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1387 netif_tx_wake_queue(txq); 1388 1389 return 0; 1390 } 1391 1392 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 1393 { 1394 struct virtio_net_hdr_mrg_rxbuf *hdr; 1395 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 1396 struct virtnet_info *vi = sq->vq->vdev->priv; 1397 int num_sg; 1398 unsigned hdr_len = vi->hdr_len; 1399 bool can_push; 1400 1401 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 1402 1403 can_push = vi->any_header_sg && 1404 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 1405 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 1406 /* Even if we can, don't push here yet as this would skew 1407 * csum_start offset below. */ 1408 if (can_push) 1409 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 1410 else 1411 hdr = skb_vnet_hdr(skb); 1412 1413 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 1414 virtio_is_little_endian(vi->vdev), false)) 1415 BUG(); 1416 1417 if (vi->mergeable_rx_bufs) 1418 hdr->num_buffers = 0; 1419 1420 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 1421 if (can_push) { 1422 __skb_push(skb, hdr_len); 1423 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 1424 if (unlikely(num_sg < 0)) 1425 return num_sg; 1426 /* Pull header back to avoid skew in tx bytes calculations. */ 1427 __skb_pull(skb, hdr_len); 1428 } else { 1429 sg_set_buf(sq->sg, hdr, hdr_len); 1430 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 1431 if (unlikely(num_sg < 0)) 1432 return num_sg; 1433 num_sg++; 1434 } 1435 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 1436 } 1437 1438 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 1439 { 1440 struct virtnet_info *vi = netdev_priv(dev); 1441 int qnum = skb_get_queue_mapping(skb); 1442 struct send_queue *sq = &vi->sq[qnum]; 1443 int err; 1444 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 1445 bool kick = !skb->xmit_more; 1446 bool use_napi = sq->napi.weight; 1447 1448 /* Free up any pending old buffers before queueing new ones. */ 1449 free_old_xmit_skbs(sq); 1450 1451 if (use_napi && kick) 1452 virtqueue_enable_cb_delayed(sq->vq); 1453 1454 /* timestamp packet in software */ 1455 skb_tx_timestamp(skb); 1456 1457 /* Try to transmit */ 1458 err = xmit_skb(sq, skb); 1459 1460 /* This should not happen! */ 1461 if (unlikely(err)) { 1462 dev->stats.tx_fifo_errors++; 1463 if (net_ratelimit()) 1464 dev_warn(&dev->dev, 1465 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err); 1466 dev->stats.tx_dropped++; 1467 dev_kfree_skb_any(skb); 1468 return NETDEV_TX_OK; 1469 } 1470 1471 /* Don't wait up for transmitted skbs to be freed. */ 1472 if (!use_napi) { 1473 skb_orphan(skb); 1474 nf_reset(skb); 1475 } 1476 1477 /* If running out of space, stop queue to avoid getting packets that we 1478 * are then unable to transmit. 1479 * An alternative would be to force queuing layer to requeue the skb by 1480 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be 1481 * returned in a normal path of operation: it means that driver is not 1482 * maintaining the TX queue stop/start state properly, and causes 1483 * the stack to do a non-trivial amount of useless work. 1484 * Since most packets only take 1 or 2 ring slots, stopping the queue 1485 * early means 16 slots are typically wasted. 1486 */ 1487 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { 1488 netif_stop_subqueue(dev, qnum); 1489 if (!use_napi && 1490 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { 1491 /* More just got used, free them then recheck. */ 1492 free_old_xmit_skbs(sq); 1493 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { 1494 netif_start_subqueue(dev, qnum); 1495 virtqueue_disable_cb(sq->vq); 1496 } 1497 } 1498 } 1499 1500 if (kick || netif_xmit_stopped(txq)) 1501 virtqueue_kick(sq->vq); 1502 1503 return NETDEV_TX_OK; 1504 } 1505 1506 /* 1507 * Send command via the control virtqueue and check status. Commands 1508 * supported by the hypervisor, as indicated by feature bits, should 1509 * never fail unless improperly formatted. 1510 */ 1511 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 1512 struct scatterlist *out) 1513 { 1514 struct scatterlist *sgs[4], hdr, stat; 1515 unsigned out_num = 0, tmp; 1516 1517 /* Caller should know better */ 1518 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 1519 1520 vi->ctrl->status = ~0; 1521 vi->ctrl->hdr.class = class; 1522 vi->ctrl->hdr.cmd = cmd; 1523 /* Add header */ 1524 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr)); 1525 sgs[out_num++] = &hdr; 1526 1527 if (out) 1528 sgs[out_num++] = out; 1529 1530 /* Add return status. */ 1531 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status)); 1532 sgs[out_num] = &stat; 1533 1534 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 1535 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 1536 1537 if (unlikely(!virtqueue_kick(vi->cvq))) 1538 return vi->ctrl->status == VIRTIO_NET_OK; 1539 1540 /* Spin for a response, the kick causes an ioport write, trapping 1541 * into the hypervisor, so the request should be handled immediately. 1542 */ 1543 while (!virtqueue_get_buf(vi->cvq, &tmp) && 1544 !virtqueue_is_broken(vi->cvq)) 1545 cpu_relax(); 1546 1547 return vi->ctrl->status == VIRTIO_NET_OK; 1548 } 1549 1550 static int virtnet_set_mac_address(struct net_device *dev, void *p) 1551 { 1552 struct virtnet_info *vi = netdev_priv(dev); 1553 struct virtio_device *vdev = vi->vdev; 1554 int ret; 1555 struct sockaddr *addr; 1556 struct scatterlist sg; 1557 1558 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 1559 return -EOPNOTSUPP; 1560 1561 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 1562 if (!addr) 1563 return -ENOMEM; 1564 1565 ret = eth_prepare_mac_addr_change(dev, addr); 1566 if (ret) 1567 goto out; 1568 1569 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 1570 sg_init_one(&sg, addr->sa_data, dev->addr_len); 1571 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1572 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 1573 dev_warn(&vdev->dev, 1574 "Failed to set mac address by vq command.\n"); 1575 ret = -EINVAL; 1576 goto out; 1577 } 1578 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 1579 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 1580 unsigned int i; 1581 1582 /* Naturally, this has an atomicity problem. */ 1583 for (i = 0; i < dev->addr_len; i++) 1584 virtio_cwrite8(vdev, 1585 offsetof(struct virtio_net_config, mac) + 1586 i, addr->sa_data[i]); 1587 } 1588 1589 eth_commit_mac_addr_change(dev, p); 1590 ret = 0; 1591 1592 out: 1593 kfree(addr); 1594 return ret; 1595 } 1596 1597 static void virtnet_stats(struct net_device *dev, 1598 struct rtnl_link_stats64 *tot) 1599 { 1600 struct virtnet_info *vi = netdev_priv(dev); 1601 unsigned int start; 1602 int i; 1603 1604 for (i = 0; i < vi->max_queue_pairs; i++) { 1605 u64 tpackets, tbytes, rpackets, rbytes; 1606 struct receive_queue *rq = &vi->rq[i]; 1607 struct send_queue *sq = &vi->sq[i]; 1608 1609 do { 1610 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1611 tpackets = sq->stats.packets; 1612 tbytes = sq->stats.bytes; 1613 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1614 1615 do { 1616 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1617 rpackets = rq->stats.packets; 1618 rbytes = rq->stats.bytes; 1619 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 1620 1621 tot->rx_packets += rpackets; 1622 tot->tx_packets += tpackets; 1623 tot->rx_bytes += rbytes; 1624 tot->tx_bytes += tbytes; 1625 } 1626 1627 tot->tx_dropped = dev->stats.tx_dropped; 1628 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 1629 tot->rx_dropped = dev->stats.rx_dropped; 1630 tot->rx_length_errors = dev->stats.rx_length_errors; 1631 tot->rx_frame_errors = dev->stats.rx_frame_errors; 1632 } 1633 1634 #ifdef CONFIG_NET_POLL_CONTROLLER 1635 static void virtnet_netpoll(struct net_device *dev) 1636 { 1637 struct virtnet_info *vi = netdev_priv(dev); 1638 int i; 1639 1640 for (i = 0; i < vi->curr_queue_pairs; i++) 1641 napi_schedule(&vi->rq[i].napi); 1642 } 1643 #endif 1644 1645 static void virtnet_ack_link_announce(struct virtnet_info *vi) 1646 { 1647 rtnl_lock(); 1648 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 1649 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 1650 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 1651 rtnl_unlock(); 1652 } 1653 1654 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1655 { 1656 struct scatterlist sg; 1657 struct net_device *dev = vi->dev; 1658 1659 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 1660 return 0; 1661 1662 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 1663 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq)); 1664 1665 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 1666 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 1667 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 1668 queue_pairs); 1669 return -EINVAL; 1670 } else { 1671 vi->curr_queue_pairs = queue_pairs; 1672 /* virtnet_open() will refill when device is going to up. */ 1673 if (dev->flags & IFF_UP) 1674 schedule_delayed_work(&vi->refill, 0); 1675 } 1676 1677 return 0; 1678 } 1679 1680 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1681 { 1682 int err; 1683 1684 rtnl_lock(); 1685 err = _virtnet_set_queues(vi, queue_pairs); 1686 rtnl_unlock(); 1687 return err; 1688 } 1689 1690 static int virtnet_close(struct net_device *dev) 1691 { 1692 struct virtnet_info *vi = netdev_priv(dev); 1693 int i; 1694 1695 /* Make sure refill_work doesn't re-enable napi! */ 1696 cancel_delayed_work_sync(&vi->refill); 1697 1698 for (i = 0; i < vi->max_queue_pairs; i++) { 1699 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1700 napi_disable(&vi->rq[i].napi); 1701 virtnet_napi_tx_disable(&vi->sq[i].napi); 1702 } 1703 1704 return 0; 1705 } 1706 1707 static void virtnet_set_rx_mode(struct net_device *dev) 1708 { 1709 struct virtnet_info *vi = netdev_priv(dev); 1710 struct scatterlist sg[2]; 1711 struct virtio_net_ctrl_mac *mac_data; 1712 struct netdev_hw_addr *ha; 1713 int uc_count; 1714 int mc_count; 1715 void *buf; 1716 int i; 1717 1718 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 1719 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 1720 return; 1721 1722 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0); 1723 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 1724 1725 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc)); 1726 1727 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1728 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 1729 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 1730 vi->ctrl->promisc ? "en" : "dis"); 1731 1732 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti)); 1733 1734 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1735 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 1736 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 1737 vi->ctrl->allmulti ? "en" : "dis"); 1738 1739 uc_count = netdev_uc_count(dev); 1740 mc_count = netdev_mc_count(dev); 1741 /* MAC filter - use one buffer for both lists */ 1742 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 1743 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 1744 mac_data = buf; 1745 if (!buf) 1746 return; 1747 1748 sg_init_table(sg, 2); 1749 1750 /* Store the unicast list and count in the front of the buffer */ 1751 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 1752 i = 0; 1753 netdev_for_each_uc_addr(ha, dev) 1754 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1755 1756 sg_set_buf(&sg[0], mac_data, 1757 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 1758 1759 /* multicast list and count fill the end */ 1760 mac_data = (void *)&mac_data->macs[uc_count][0]; 1761 1762 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 1763 i = 0; 1764 netdev_for_each_mc_addr(ha, dev) 1765 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1766 1767 sg_set_buf(&sg[1], mac_data, 1768 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 1769 1770 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1771 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 1772 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 1773 1774 kfree(buf); 1775 } 1776 1777 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 1778 __be16 proto, u16 vid) 1779 { 1780 struct virtnet_info *vi = netdev_priv(dev); 1781 struct scatterlist sg; 1782 1783 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 1784 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 1785 1786 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1787 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 1788 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 1789 return 0; 1790 } 1791 1792 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 1793 __be16 proto, u16 vid) 1794 { 1795 struct virtnet_info *vi = netdev_priv(dev); 1796 struct scatterlist sg; 1797 1798 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 1799 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 1800 1801 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1802 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 1803 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 1804 return 0; 1805 } 1806 1807 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu) 1808 { 1809 int i; 1810 1811 if (vi->affinity_hint_set) { 1812 for (i = 0; i < vi->max_queue_pairs; i++) { 1813 virtqueue_set_affinity(vi->rq[i].vq, -1); 1814 virtqueue_set_affinity(vi->sq[i].vq, -1); 1815 } 1816 1817 vi->affinity_hint_set = false; 1818 } 1819 } 1820 1821 static void virtnet_set_affinity(struct virtnet_info *vi) 1822 { 1823 int i; 1824 int cpu; 1825 1826 /* In multiqueue mode, when the number of cpu is equal to the number of 1827 * queue pairs, we let the queue pairs to be private to one cpu by 1828 * setting the affinity hint to eliminate the contention. 1829 */ 1830 if (vi->curr_queue_pairs == 1 || 1831 vi->max_queue_pairs != num_online_cpus()) { 1832 virtnet_clean_affinity(vi, -1); 1833 return; 1834 } 1835 1836 i = 0; 1837 for_each_online_cpu(cpu) { 1838 virtqueue_set_affinity(vi->rq[i].vq, cpu); 1839 virtqueue_set_affinity(vi->sq[i].vq, cpu); 1840 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i); 1841 i++; 1842 } 1843 1844 vi->affinity_hint_set = true; 1845 } 1846 1847 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 1848 { 1849 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1850 node); 1851 virtnet_set_affinity(vi); 1852 return 0; 1853 } 1854 1855 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 1856 { 1857 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1858 node_dead); 1859 virtnet_set_affinity(vi); 1860 return 0; 1861 } 1862 1863 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 1864 { 1865 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1866 node); 1867 1868 virtnet_clean_affinity(vi, cpu); 1869 return 0; 1870 } 1871 1872 static enum cpuhp_state virtionet_online; 1873 1874 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 1875 { 1876 int ret; 1877 1878 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 1879 if (ret) 1880 return ret; 1881 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1882 &vi->node_dead); 1883 if (!ret) 1884 return ret; 1885 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1886 return ret; 1887 } 1888 1889 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 1890 { 1891 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1892 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1893 &vi->node_dead); 1894 } 1895 1896 static void virtnet_get_ringparam(struct net_device *dev, 1897 struct ethtool_ringparam *ring) 1898 { 1899 struct virtnet_info *vi = netdev_priv(dev); 1900 1901 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq); 1902 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq); 1903 ring->rx_pending = ring->rx_max_pending; 1904 ring->tx_pending = ring->tx_max_pending; 1905 } 1906 1907 1908 static void virtnet_get_drvinfo(struct net_device *dev, 1909 struct ethtool_drvinfo *info) 1910 { 1911 struct virtnet_info *vi = netdev_priv(dev); 1912 struct virtio_device *vdev = vi->vdev; 1913 1914 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 1915 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 1916 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 1917 1918 } 1919 1920 /* TODO: Eliminate OOO packets during switching */ 1921 static int virtnet_set_channels(struct net_device *dev, 1922 struct ethtool_channels *channels) 1923 { 1924 struct virtnet_info *vi = netdev_priv(dev); 1925 u16 queue_pairs = channels->combined_count; 1926 int err; 1927 1928 /* We don't support separate rx/tx channels. 1929 * We don't allow setting 'other' channels. 1930 */ 1931 if (channels->rx_count || channels->tx_count || channels->other_count) 1932 return -EINVAL; 1933 1934 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 1935 return -EINVAL; 1936 1937 /* For now we don't support modifying channels while XDP is loaded 1938 * also when XDP is loaded all RX queues have XDP programs so we only 1939 * need to check a single RX queue. 1940 */ 1941 if (vi->rq[0].xdp_prog) 1942 return -EINVAL; 1943 1944 get_online_cpus(); 1945 err = _virtnet_set_queues(vi, queue_pairs); 1946 if (!err) { 1947 netif_set_real_num_tx_queues(dev, queue_pairs); 1948 netif_set_real_num_rx_queues(dev, queue_pairs); 1949 1950 virtnet_set_affinity(vi); 1951 } 1952 put_online_cpus(); 1953 1954 return err; 1955 } 1956 1957 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 1958 { 1959 struct virtnet_info *vi = netdev_priv(dev); 1960 char *p = (char *)data; 1961 unsigned int i, j; 1962 1963 switch (stringset) { 1964 case ETH_SS_STATS: 1965 for (i = 0; i < vi->curr_queue_pairs; i++) { 1966 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 1967 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s", 1968 i, virtnet_rq_stats_desc[j].desc); 1969 p += ETH_GSTRING_LEN; 1970 } 1971 } 1972 1973 for (i = 0; i < vi->curr_queue_pairs; i++) { 1974 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 1975 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s", 1976 i, virtnet_sq_stats_desc[j].desc); 1977 p += ETH_GSTRING_LEN; 1978 } 1979 } 1980 break; 1981 } 1982 } 1983 1984 static int virtnet_get_sset_count(struct net_device *dev, int sset) 1985 { 1986 struct virtnet_info *vi = netdev_priv(dev); 1987 1988 switch (sset) { 1989 case ETH_SS_STATS: 1990 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 1991 VIRTNET_SQ_STATS_LEN); 1992 default: 1993 return -EOPNOTSUPP; 1994 } 1995 } 1996 1997 static void virtnet_get_ethtool_stats(struct net_device *dev, 1998 struct ethtool_stats *stats, u64 *data) 1999 { 2000 struct virtnet_info *vi = netdev_priv(dev); 2001 unsigned int idx = 0, start, i, j; 2002 const u8 *stats_base; 2003 size_t offset; 2004 2005 for (i = 0; i < vi->curr_queue_pairs; i++) { 2006 struct receive_queue *rq = &vi->rq[i]; 2007 2008 stats_base = (u8 *)&rq->stats; 2009 do { 2010 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 2011 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 2012 offset = virtnet_rq_stats_desc[j].offset; 2013 data[idx + j] = *(u64 *)(stats_base + offset); 2014 } 2015 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 2016 idx += VIRTNET_RQ_STATS_LEN; 2017 } 2018 2019 for (i = 0; i < vi->curr_queue_pairs; i++) { 2020 struct send_queue *sq = &vi->sq[i]; 2021 2022 stats_base = (u8 *)&sq->stats; 2023 do { 2024 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 2025 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 2026 offset = virtnet_sq_stats_desc[j].offset; 2027 data[idx + j] = *(u64 *)(stats_base + offset); 2028 } 2029 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 2030 idx += VIRTNET_SQ_STATS_LEN; 2031 } 2032 } 2033 2034 static void virtnet_get_channels(struct net_device *dev, 2035 struct ethtool_channels *channels) 2036 { 2037 struct virtnet_info *vi = netdev_priv(dev); 2038 2039 channels->combined_count = vi->curr_queue_pairs; 2040 channels->max_combined = vi->max_queue_pairs; 2041 channels->max_other = 0; 2042 channels->rx_count = 0; 2043 channels->tx_count = 0; 2044 channels->other_count = 0; 2045 } 2046 2047 /* Check if the user is trying to change anything besides speed/duplex */ 2048 static bool 2049 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd) 2050 { 2051 struct ethtool_link_ksettings diff1 = *cmd; 2052 struct ethtool_link_ksettings diff2 = {}; 2053 2054 /* cmd is always set so we need to clear it, validate the port type 2055 * and also without autonegotiation we can ignore advertising 2056 */ 2057 diff1.base.speed = 0; 2058 diff2.base.port = PORT_OTHER; 2059 ethtool_link_ksettings_zero_link_mode(&diff1, advertising); 2060 diff1.base.duplex = 0; 2061 diff1.base.cmd = 0; 2062 diff1.base.link_mode_masks_nwords = 0; 2063 2064 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) && 2065 bitmap_empty(diff1.link_modes.supported, 2066 __ETHTOOL_LINK_MODE_MASK_NBITS) && 2067 bitmap_empty(diff1.link_modes.advertising, 2068 __ETHTOOL_LINK_MODE_MASK_NBITS) && 2069 bitmap_empty(diff1.link_modes.lp_advertising, 2070 __ETHTOOL_LINK_MODE_MASK_NBITS); 2071 } 2072 2073 static int virtnet_set_link_ksettings(struct net_device *dev, 2074 const struct ethtool_link_ksettings *cmd) 2075 { 2076 struct virtnet_info *vi = netdev_priv(dev); 2077 u32 speed; 2078 2079 speed = cmd->base.speed; 2080 /* don't allow custom speed and duplex */ 2081 if (!ethtool_validate_speed(speed) || 2082 !ethtool_validate_duplex(cmd->base.duplex) || 2083 !virtnet_validate_ethtool_cmd(cmd)) 2084 return -EINVAL; 2085 vi->speed = speed; 2086 vi->duplex = cmd->base.duplex; 2087 2088 return 0; 2089 } 2090 2091 static int virtnet_get_link_ksettings(struct net_device *dev, 2092 struct ethtool_link_ksettings *cmd) 2093 { 2094 struct virtnet_info *vi = netdev_priv(dev); 2095 2096 cmd->base.speed = vi->speed; 2097 cmd->base.duplex = vi->duplex; 2098 cmd->base.port = PORT_OTHER; 2099 2100 return 0; 2101 } 2102 2103 static void virtnet_init_settings(struct net_device *dev) 2104 { 2105 struct virtnet_info *vi = netdev_priv(dev); 2106 2107 vi->speed = SPEED_UNKNOWN; 2108 vi->duplex = DUPLEX_UNKNOWN; 2109 } 2110 2111 static void virtnet_update_settings(struct virtnet_info *vi) 2112 { 2113 u32 speed; 2114 u8 duplex; 2115 2116 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 2117 return; 2118 2119 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config, 2120 speed)); 2121 if (ethtool_validate_speed(speed)) 2122 vi->speed = speed; 2123 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config, 2124 duplex)); 2125 if (ethtool_validate_duplex(duplex)) 2126 vi->duplex = duplex; 2127 } 2128 2129 static const struct ethtool_ops virtnet_ethtool_ops = { 2130 .get_drvinfo = virtnet_get_drvinfo, 2131 .get_link = ethtool_op_get_link, 2132 .get_ringparam = virtnet_get_ringparam, 2133 .get_strings = virtnet_get_strings, 2134 .get_sset_count = virtnet_get_sset_count, 2135 .get_ethtool_stats = virtnet_get_ethtool_stats, 2136 .set_channels = virtnet_set_channels, 2137 .get_channels = virtnet_get_channels, 2138 .get_ts_info = ethtool_op_get_ts_info, 2139 .get_link_ksettings = virtnet_get_link_ksettings, 2140 .set_link_ksettings = virtnet_set_link_ksettings, 2141 }; 2142 2143 static void virtnet_freeze_down(struct virtio_device *vdev) 2144 { 2145 struct virtnet_info *vi = vdev->priv; 2146 int i; 2147 2148 /* Make sure no work handler is accessing the device */ 2149 flush_work(&vi->config_work); 2150 2151 netif_device_detach(vi->dev); 2152 netif_tx_disable(vi->dev); 2153 cancel_delayed_work_sync(&vi->refill); 2154 2155 if (netif_running(vi->dev)) { 2156 for (i = 0; i < vi->max_queue_pairs; i++) { 2157 napi_disable(&vi->rq[i].napi); 2158 virtnet_napi_tx_disable(&vi->sq[i].napi); 2159 } 2160 } 2161 } 2162 2163 static int init_vqs(struct virtnet_info *vi); 2164 2165 static int virtnet_restore_up(struct virtio_device *vdev) 2166 { 2167 struct virtnet_info *vi = vdev->priv; 2168 int err, i; 2169 2170 err = init_vqs(vi); 2171 if (err) 2172 return err; 2173 2174 virtio_device_ready(vdev); 2175 2176 if (netif_running(vi->dev)) { 2177 for (i = 0; i < vi->curr_queue_pairs; i++) 2178 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 2179 schedule_delayed_work(&vi->refill, 0); 2180 2181 for (i = 0; i < vi->max_queue_pairs; i++) { 2182 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2183 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 2184 &vi->sq[i].napi); 2185 } 2186 } 2187 2188 netif_device_attach(vi->dev); 2189 return err; 2190 } 2191 2192 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 2193 { 2194 struct scatterlist sg; 2195 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads); 2196 2197 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads)); 2198 2199 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 2200 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 2201 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n"); 2202 return -EINVAL; 2203 } 2204 2205 return 0; 2206 } 2207 2208 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 2209 { 2210 u64 offloads = 0; 2211 2212 if (!vi->guest_offloads) 2213 return 0; 2214 2215 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2216 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2217 2218 return virtnet_set_guest_offloads(vi, offloads); 2219 } 2220 2221 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 2222 { 2223 u64 offloads = vi->guest_offloads; 2224 2225 if (!vi->guest_offloads) 2226 return 0; 2227 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2228 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2229 2230 return virtnet_set_guest_offloads(vi, offloads); 2231 } 2232 2233 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 2234 struct netlink_ext_ack *extack) 2235 { 2236 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr); 2237 struct virtnet_info *vi = netdev_priv(dev); 2238 struct bpf_prog *old_prog; 2239 u16 xdp_qp = 0, curr_qp; 2240 int i, err; 2241 2242 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 2243 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 2244 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 2245 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 2246 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) { 2247 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first"); 2248 return -EOPNOTSUPP; 2249 } 2250 2251 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 2252 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 2253 return -EINVAL; 2254 } 2255 2256 if (dev->mtu > max_sz) { 2257 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP"); 2258 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz); 2259 return -EINVAL; 2260 } 2261 2262 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 2263 if (prog) 2264 xdp_qp = nr_cpu_ids; 2265 2266 /* XDP requires extra queues for XDP_TX */ 2267 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 2268 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available"); 2269 netdev_warn(dev, "request %i queues but max is %i\n", 2270 curr_qp + xdp_qp, vi->max_queue_pairs); 2271 return -ENOMEM; 2272 } 2273 2274 if (prog) { 2275 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1); 2276 if (IS_ERR(prog)) 2277 return PTR_ERR(prog); 2278 } 2279 2280 /* Make sure NAPI is not using any XDP TX queues for RX. */ 2281 if (netif_running(dev)) 2282 for (i = 0; i < vi->max_queue_pairs; i++) 2283 napi_disable(&vi->rq[i].napi); 2284 2285 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 2286 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 2287 if (err) 2288 goto err; 2289 vi->xdp_queue_pairs = xdp_qp; 2290 2291 for (i = 0; i < vi->max_queue_pairs; i++) { 2292 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2293 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 2294 if (i == 0) { 2295 if (!old_prog) 2296 virtnet_clear_guest_offloads(vi); 2297 if (!prog) 2298 virtnet_restore_guest_offloads(vi); 2299 } 2300 if (old_prog) 2301 bpf_prog_put(old_prog); 2302 if (netif_running(dev)) 2303 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2304 } 2305 2306 return 0; 2307 2308 err: 2309 for (i = 0; i < vi->max_queue_pairs; i++) 2310 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2311 if (prog) 2312 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 2313 return err; 2314 } 2315 2316 static u32 virtnet_xdp_query(struct net_device *dev) 2317 { 2318 struct virtnet_info *vi = netdev_priv(dev); 2319 const struct bpf_prog *xdp_prog; 2320 int i; 2321 2322 for (i = 0; i < vi->max_queue_pairs; i++) { 2323 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2324 if (xdp_prog) 2325 return xdp_prog->aux->id; 2326 } 2327 return 0; 2328 } 2329 2330 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 2331 { 2332 switch (xdp->command) { 2333 case XDP_SETUP_PROG: 2334 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 2335 case XDP_QUERY_PROG: 2336 xdp->prog_id = virtnet_xdp_query(dev); 2337 xdp->prog_attached = !!xdp->prog_id; 2338 return 0; 2339 default: 2340 return -EINVAL; 2341 } 2342 } 2343 2344 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf, 2345 size_t len) 2346 { 2347 struct virtnet_info *vi = netdev_priv(dev); 2348 int ret; 2349 2350 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 2351 return -EOPNOTSUPP; 2352 2353 ret = snprintf(buf, len, "sby"); 2354 if (ret >= len) 2355 return -EOPNOTSUPP; 2356 2357 return 0; 2358 } 2359 2360 static const struct net_device_ops virtnet_netdev = { 2361 .ndo_open = virtnet_open, 2362 .ndo_stop = virtnet_close, 2363 .ndo_start_xmit = start_xmit, 2364 .ndo_validate_addr = eth_validate_addr, 2365 .ndo_set_mac_address = virtnet_set_mac_address, 2366 .ndo_set_rx_mode = virtnet_set_rx_mode, 2367 .ndo_get_stats64 = virtnet_stats, 2368 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 2369 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 2370 #ifdef CONFIG_NET_POLL_CONTROLLER 2371 .ndo_poll_controller = virtnet_netpoll, 2372 #endif 2373 .ndo_bpf = virtnet_xdp, 2374 .ndo_xdp_xmit = virtnet_xdp_xmit, 2375 .ndo_features_check = passthru_features_check, 2376 .ndo_get_phys_port_name = virtnet_get_phys_port_name, 2377 }; 2378 2379 static void virtnet_config_changed_work(struct work_struct *work) 2380 { 2381 struct virtnet_info *vi = 2382 container_of(work, struct virtnet_info, config_work); 2383 u16 v; 2384 2385 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 2386 struct virtio_net_config, status, &v) < 0) 2387 return; 2388 2389 if (v & VIRTIO_NET_S_ANNOUNCE) { 2390 netdev_notify_peers(vi->dev); 2391 virtnet_ack_link_announce(vi); 2392 } 2393 2394 /* Ignore unknown (future) status bits */ 2395 v &= VIRTIO_NET_S_LINK_UP; 2396 2397 if (vi->status == v) 2398 return; 2399 2400 vi->status = v; 2401 2402 if (vi->status & VIRTIO_NET_S_LINK_UP) { 2403 virtnet_update_settings(vi); 2404 netif_carrier_on(vi->dev); 2405 netif_tx_wake_all_queues(vi->dev); 2406 } else { 2407 netif_carrier_off(vi->dev); 2408 netif_tx_stop_all_queues(vi->dev); 2409 } 2410 } 2411 2412 static void virtnet_config_changed(struct virtio_device *vdev) 2413 { 2414 struct virtnet_info *vi = vdev->priv; 2415 2416 schedule_work(&vi->config_work); 2417 } 2418 2419 static void virtnet_free_queues(struct virtnet_info *vi) 2420 { 2421 int i; 2422 2423 for (i = 0; i < vi->max_queue_pairs; i++) { 2424 napi_hash_del(&vi->rq[i].napi); 2425 netif_napi_del(&vi->rq[i].napi); 2426 netif_napi_del(&vi->sq[i].napi); 2427 } 2428 2429 /* We called napi_hash_del() before netif_napi_del(), 2430 * we need to respect an RCU grace period before freeing vi->rq 2431 */ 2432 synchronize_net(); 2433 2434 kfree(vi->rq); 2435 kfree(vi->sq); 2436 kfree(vi->ctrl); 2437 } 2438 2439 static void _free_receive_bufs(struct virtnet_info *vi) 2440 { 2441 struct bpf_prog *old_prog; 2442 int i; 2443 2444 for (i = 0; i < vi->max_queue_pairs; i++) { 2445 while (vi->rq[i].pages) 2446 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 2447 2448 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2449 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 2450 if (old_prog) 2451 bpf_prog_put(old_prog); 2452 } 2453 } 2454 2455 static void free_receive_bufs(struct virtnet_info *vi) 2456 { 2457 rtnl_lock(); 2458 _free_receive_bufs(vi); 2459 rtnl_unlock(); 2460 } 2461 2462 static void free_receive_page_frags(struct virtnet_info *vi) 2463 { 2464 int i; 2465 for (i = 0; i < vi->max_queue_pairs; i++) 2466 if (vi->rq[i].alloc_frag.page) 2467 put_page(vi->rq[i].alloc_frag.page); 2468 } 2469 2470 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) 2471 { 2472 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs)) 2473 return false; 2474 else if (q < vi->curr_queue_pairs) 2475 return true; 2476 else 2477 return false; 2478 } 2479 2480 static void free_unused_bufs(struct virtnet_info *vi) 2481 { 2482 void *buf; 2483 int i; 2484 2485 for (i = 0; i < vi->max_queue_pairs; i++) { 2486 struct virtqueue *vq = vi->sq[i].vq; 2487 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2488 if (!is_xdp_raw_buffer_queue(vi, i)) 2489 dev_kfree_skb(buf); 2490 else 2491 put_page(virt_to_head_page(buf)); 2492 } 2493 } 2494 2495 for (i = 0; i < vi->max_queue_pairs; i++) { 2496 struct virtqueue *vq = vi->rq[i].vq; 2497 2498 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2499 if (vi->mergeable_rx_bufs) { 2500 put_page(virt_to_head_page(buf)); 2501 } else if (vi->big_packets) { 2502 give_pages(&vi->rq[i], buf); 2503 } else { 2504 put_page(virt_to_head_page(buf)); 2505 } 2506 } 2507 } 2508 } 2509 2510 static void virtnet_del_vqs(struct virtnet_info *vi) 2511 { 2512 struct virtio_device *vdev = vi->vdev; 2513 2514 virtnet_clean_affinity(vi, -1); 2515 2516 vdev->config->del_vqs(vdev); 2517 2518 virtnet_free_queues(vi); 2519 } 2520 2521 /* How large should a single buffer be so a queue full of these can fit at 2522 * least one full packet? 2523 * Logic below assumes the mergeable buffer header is used. 2524 */ 2525 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 2526 { 2527 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2528 unsigned int rq_size = virtqueue_get_vring_size(vq); 2529 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 2530 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 2531 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 2532 2533 return max(max(min_buf_len, hdr_len) - hdr_len, 2534 (unsigned int)GOOD_PACKET_LEN); 2535 } 2536 2537 static int virtnet_find_vqs(struct virtnet_info *vi) 2538 { 2539 vq_callback_t **callbacks; 2540 struct virtqueue **vqs; 2541 int ret = -ENOMEM; 2542 int i, total_vqs; 2543 const char **names; 2544 bool *ctx; 2545 2546 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 2547 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 2548 * possible control vq. 2549 */ 2550 total_vqs = vi->max_queue_pairs * 2 + 2551 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 2552 2553 /* Allocate space for find_vqs parameters */ 2554 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL); 2555 if (!vqs) 2556 goto err_vq; 2557 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL); 2558 if (!callbacks) 2559 goto err_callback; 2560 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL); 2561 if (!names) 2562 goto err_names; 2563 if (!vi->big_packets || vi->mergeable_rx_bufs) { 2564 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL); 2565 if (!ctx) 2566 goto err_ctx; 2567 } else { 2568 ctx = NULL; 2569 } 2570 2571 /* Parameters for control virtqueue, if any */ 2572 if (vi->has_cvq) { 2573 callbacks[total_vqs - 1] = NULL; 2574 names[total_vqs - 1] = "control"; 2575 } 2576 2577 /* Allocate/initialize parameters for send/receive virtqueues */ 2578 for (i = 0; i < vi->max_queue_pairs; i++) { 2579 callbacks[rxq2vq(i)] = skb_recv_done; 2580 callbacks[txq2vq(i)] = skb_xmit_done; 2581 sprintf(vi->rq[i].name, "input.%d", i); 2582 sprintf(vi->sq[i].name, "output.%d", i); 2583 names[rxq2vq(i)] = vi->rq[i].name; 2584 names[txq2vq(i)] = vi->sq[i].name; 2585 if (ctx) 2586 ctx[rxq2vq(i)] = true; 2587 } 2588 2589 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks, 2590 names, ctx, NULL); 2591 if (ret) 2592 goto err_find; 2593 2594 if (vi->has_cvq) { 2595 vi->cvq = vqs[total_vqs - 1]; 2596 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 2597 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 2598 } 2599 2600 for (i = 0; i < vi->max_queue_pairs; i++) { 2601 vi->rq[i].vq = vqs[rxq2vq(i)]; 2602 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 2603 vi->sq[i].vq = vqs[txq2vq(i)]; 2604 } 2605 2606 /* run here: ret == 0. */ 2607 2608 2609 err_find: 2610 kfree(ctx); 2611 err_ctx: 2612 kfree(names); 2613 err_names: 2614 kfree(callbacks); 2615 err_callback: 2616 kfree(vqs); 2617 err_vq: 2618 return ret; 2619 } 2620 2621 static int virtnet_alloc_queues(struct virtnet_info *vi) 2622 { 2623 int i; 2624 2625 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL); 2626 if (!vi->ctrl) 2627 goto err_ctrl; 2628 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL); 2629 if (!vi->sq) 2630 goto err_sq; 2631 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL); 2632 if (!vi->rq) 2633 goto err_rq; 2634 2635 INIT_DELAYED_WORK(&vi->refill, refill_work); 2636 for (i = 0; i < vi->max_queue_pairs; i++) { 2637 vi->rq[i].pages = NULL; 2638 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll, 2639 napi_weight); 2640 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx, 2641 napi_tx ? napi_weight : 0); 2642 2643 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 2644 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 2645 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 2646 2647 u64_stats_init(&vi->rq[i].stats.syncp); 2648 u64_stats_init(&vi->sq[i].stats.syncp); 2649 } 2650 2651 return 0; 2652 2653 err_rq: 2654 kfree(vi->sq); 2655 err_sq: 2656 kfree(vi->ctrl); 2657 err_ctrl: 2658 return -ENOMEM; 2659 } 2660 2661 static int init_vqs(struct virtnet_info *vi) 2662 { 2663 int ret; 2664 2665 /* Allocate send & receive queues */ 2666 ret = virtnet_alloc_queues(vi); 2667 if (ret) 2668 goto err; 2669 2670 ret = virtnet_find_vqs(vi); 2671 if (ret) 2672 goto err_free; 2673 2674 get_online_cpus(); 2675 virtnet_set_affinity(vi); 2676 put_online_cpus(); 2677 2678 return 0; 2679 2680 err_free: 2681 virtnet_free_queues(vi); 2682 err: 2683 return ret; 2684 } 2685 2686 #ifdef CONFIG_SYSFS 2687 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 2688 char *buf) 2689 { 2690 struct virtnet_info *vi = netdev_priv(queue->dev); 2691 unsigned int queue_index = get_netdev_rx_queue_index(queue); 2692 unsigned int headroom = virtnet_get_headroom(vi); 2693 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 2694 struct ewma_pkt_len *avg; 2695 2696 BUG_ON(queue_index >= vi->max_queue_pairs); 2697 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 2698 return sprintf(buf, "%u\n", 2699 get_mergeable_buf_len(&vi->rq[queue_index], avg, 2700 SKB_DATA_ALIGN(headroom + tailroom))); 2701 } 2702 2703 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 2704 __ATTR_RO(mergeable_rx_buffer_size); 2705 2706 static struct attribute *virtio_net_mrg_rx_attrs[] = { 2707 &mergeable_rx_buffer_size_attribute.attr, 2708 NULL 2709 }; 2710 2711 static const struct attribute_group virtio_net_mrg_rx_group = { 2712 .name = "virtio_net", 2713 .attrs = virtio_net_mrg_rx_attrs 2714 }; 2715 #endif 2716 2717 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 2718 unsigned int fbit, 2719 const char *fname, const char *dname) 2720 { 2721 if (!virtio_has_feature(vdev, fbit)) 2722 return false; 2723 2724 dev_err(&vdev->dev, "device advertises feature %s but not %s", 2725 fname, dname); 2726 2727 return true; 2728 } 2729 2730 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 2731 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 2732 2733 static bool virtnet_validate_features(struct virtio_device *vdev) 2734 { 2735 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 2736 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 2737 "VIRTIO_NET_F_CTRL_VQ") || 2738 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 2739 "VIRTIO_NET_F_CTRL_VQ") || 2740 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 2741 "VIRTIO_NET_F_CTRL_VQ") || 2742 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 2743 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 2744 "VIRTIO_NET_F_CTRL_VQ"))) { 2745 return false; 2746 } 2747 2748 return true; 2749 } 2750 2751 #define MIN_MTU ETH_MIN_MTU 2752 #define MAX_MTU ETH_MAX_MTU 2753 2754 static int virtnet_validate(struct virtio_device *vdev) 2755 { 2756 if (!vdev->config->get) { 2757 dev_err(&vdev->dev, "%s failure: config access disabled\n", 2758 __func__); 2759 return -EINVAL; 2760 } 2761 2762 if (!virtnet_validate_features(vdev)) 2763 return -EINVAL; 2764 2765 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2766 int mtu = virtio_cread16(vdev, 2767 offsetof(struct virtio_net_config, 2768 mtu)); 2769 if (mtu < MIN_MTU) 2770 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 2771 } 2772 2773 return 0; 2774 } 2775 2776 static int virtnet_probe(struct virtio_device *vdev) 2777 { 2778 int i, err = -ENOMEM; 2779 struct net_device *dev; 2780 struct virtnet_info *vi; 2781 u16 max_queue_pairs; 2782 int mtu; 2783 2784 /* Find if host supports multiqueue virtio_net device */ 2785 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ, 2786 struct virtio_net_config, 2787 max_virtqueue_pairs, &max_queue_pairs); 2788 2789 /* We need at least 2 queue's */ 2790 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 2791 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 2792 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2793 max_queue_pairs = 1; 2794 2795 /* Allocate ourselves a network device with room for our info */ 2796 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 2797 if (!dev) 2798 return -ENOMEM; 2799 2800 /* Set up network device as normal. */ 2801 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE; 2802 dev->netdev_ops = &virtnet_netdev; 2803 dev->features = NETIF_F_HIGHDMA; 2804 2805 dev->ethtool_ops = &virtnet_ethtool_ops; 2806 SET_NETDEV_DEV(dev, &vdev->dev); 2807 2808 /* Do we support "hardware" checksums? */ 2809 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 2810 /* This opens up the world of extra features. */ 2811 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2812 if (csum) 2813 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2814 2815 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 2816 dev->hw_features |= NETIF_F_TSO 2817 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 2818 } 2819 /* Individual feature bits: what can host handle? */ 2820 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 2821 dev->hw_features |= NETIF_F_TSO; 2822 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 2823 dev->hw_features |= NETIF_F_TSO6; 2824 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 2825 dev->hw_features |= NETIF_F_TSO_ECN; 2826 2827 dev->features |= NETIF_F_GSO_ROBUST; 2828 2829 if (gso) 2830 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 2831 /* (!csum && gso) case will be fixed by register_netdev() */ 2832 } 2833 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 2834 dev->features |= NETIF_F_RXCSUM; 2835 2836 dev->vlan_features = dev->features; 2837 2838 /* MTU range: 68 - 65535 */ 2839 dev->min_mtu = MIN_MTU; 2840 dev->max_mtu = MAX_MTU; 2841 2842 /* Configuration may specify what MAC to use. Otherwise random. */ 2843 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) 2844 virtio_cread_bytes(vdev, 2845 offsetof(struct virtio_net_config, mac), 2846 dev->dev_addr, dev->addr_len); 2847 else 2848 eth_hw_addr_random(dev); 2849 2850 /* Set up our device-specific information */ 2851 vi = netdev_priv(dev); 2852 vi->dev = dev; 2853 vi->vdev = vdev; 2854 vdev->priv = vi; 2855 2856 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 2857 2858 /* If we can receive ANY GSO packets, we must allocate large ones. */ 2859 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 2860 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) || 2861 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) || 2862 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO)) 2863 vi->big_packets = true; 2864 2865 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) 2866 vi->mergeable_rx_bufs = true; 2867 2868 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 2869 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2870 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2871 else 2872 vi->hdr_len = sizeof(struct virtio_net_hdr); 2873 2874 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 2875 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2876 vi->any_header_sg = true; 2877 2878 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2879 vi->has_cvq = true; 2880 2881 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2882 mtu = virtio_cread16(vdev, 2883 offsetof(struct virtio_net_config, 2884 mtu)); 2885 if (mtu < dev->min_mtu) { 2886 /* Should never trigger: MTU was previously validated 2887 * in virtnet_validate. 2888 */ 2889 dev_err(&vdev->dev, "device MTU appears to have changed " 2890 "it is now %d < %d", mtu, dev->min_mtu); 2891 goto free; 2892 } 2893 2894 dev->mtu = mtu; 2895 dev->max_mtu = mtu; 2896 2897 /* TODO: size buffers correctly in this case. */ 2898 if (dev->mtu > ETH_DATA_LEN) 2899 vi->big_packets = true; 2900 } 2901 2902 if (vi->any_header_sg) 2903 dev->needed_headroom = vi->hdr_len; 2904 2905 /* Enable multiqueue by default */ 2906 if (num_online_cpus() >= max_queue_pairs) 2907 vi->curr_queue_pairs = max_queue_pairs; 2908 else 2909 vi->curr_queue_pairs = num_online_cpus(); 2910 vi->max_queue_pairs = max_queue_pairs; 2911 2912 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 2913 err = init_vqs(vi); 2914 if (err) 2915 goto free; 2916 2917 #ifdef CONFIG_SYSFS 2918 if (vi->mergeable_rx_bufs) 2919 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 2920 #endif 2921 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 2922 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 2923 2924 virtnet_init_settings(dev); 2925 2926 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) { 2927 vi->failover = net_failover_create(vi->dev); 2928 if (IS_ERR(vi->failover)) { 2929 err = PTR_ERR(vi->failover); 2930 goto free_vqs; 2931 } 2932 } 2933 2934 err = register_netdev(dev); 2935 if (err) { 2936 pr_debug("virtio_net: registering device failed\n"); 2937 goto free_failover; 2938 } 2939 2940 virtio_device_ready(vdev); 2941 2942 err = virtnet_cpu_notif_add(vi); 2943 if (err) { 2944 pr_debug("virtio_net: registering cpu notifier failed\n"); 2945 goto free_unregister_netdev; 2946 } 2947 2948 virtnet_set_queues(vi, vi->curr_queue_pairs); 2949 2950 /* Assume link up if device can't report link status, 2951 otherwise get link status from config. */ 2952 netif_carrier_off(dev); 2953 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 2954 schedule_work(&vi->config_work); 2955 } else { 2956 vi->status = VIRTIO_NET_S_LINK_UP; 2957 virtnet_update_settings(vi); 2958 netif_carrier_on(dev); 2959 } 2960 2961 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 2962 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 2963 set_bit(guest_offloads[i], &vi->guest_offloads); 2964 2965 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 2966 dev->name, max_queue_pairs); 2967 2968 return 0; 2969 2970 free_unregister_netdev: 2971 vi->vdev->config->reset(vdev); 2972 2973 unregister_netdev(dev); 2974 free_failover: 2975 net_failover_destroy(vi->failover); 2976 free_vqs: 2977 cancel_delayed_work_sync(&vi->refill); 2978 free_receive_page_frags(vi); 2979 virtnet_del_vqs(vi); 2980 free: 2981 free_netdev(dev); 2982 return err; 2983 } 2984 2985 static void remove_vq_common(struct virtnet_info *vi) 2986 { 2987 vi->vdev->config->reset(vi->vdev); 2988 2989 /* Free unused buffers in both send and recv, if any. */ 2990 free_unused_bufs(vi); 2991 2992 free_receive_bufs(vi); 2993 2994 free_receive_page_frags(vi); 2995 2996 virtnet_del_vqs(vi); 2997 } 2998 2999 static void virtnet_remove(struct virtio_device *vdev) 3000 { 3001 struct virtnet_info *vi = vdev->priv; 3002 3003 virtnet_cpu_notif_remove(vi); 3004 3005 /* Make sure no work handler is accessing the device. */ 3006 flush_work(&vi->config_work); 3007 3008 unregister_netdev(vi->dev); 3009 3010 net_failover_destroy(vi->failover); 3011 3012 remove_vq_common(vi); 3013 3014 free_netdev(vi->dev); 3015 } 3016 3017 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 3018 { 3019 struct virtnet_info *vi = vdev->priv; 3020 3021 virtnet_cpu_notif_remove(vi); 3022 virtnet_freeze_down(vdev); 3023 remove_vq_common(vi); 3024 3025 return 0; 3026 } 3027 3028 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 3029 { 3030 struct virtnet_info *vi = vdev->priv; 3031 int err; 3032 3033 err = virtnet_restore_up(vdev); 3034 if (err) 3035 return err; 3036 virtnet_set_queues(vi, vi->curr_queue_pairs); 3037 3038 err = virtnet_cpu_notif_add(vi); 3039 if (err) 3040 return err; 3041 3042 return 0; 3043 } 3044 3045 static struct virtio_device_id id_table[] = { 3046 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 3047 { 0 }, 3048 }; 3049 3050 #define VIRTNET_FEATURES \ 3051 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 3052 VIRTIO_NET_F_MAC, \ 3053 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 3054 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 3055 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 3056 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 3057 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 3058 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 3059 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 3060 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 3061 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY 3062 3063 static unsigned int features[] = { 3064 VIRTNET_FEATURES, 3065 }; 3066 3067 static unsigned int features_legacy[] = { 3068 VIRTNET_FEATURES, 3069 VIRTIO_NET_F_GSO, 3070 VIRTIO_F_ANY_LAYOUT, 3071 }; 3072 3073 static struct virtio_driver virtio_net_driver = { 3074 .feature_table = features, 3075 .feature_table_size = ARRAY_SIZE(features), 3076 .feature_table_legacy = features_legacy, 3077 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 3078 .driver.name = KBUILD_MODNAME, 3079 .driver.owner = THIS_MODULE, 3080 .id_table = id_table, 3081 .validate = virtnet_validate, 3082 .probe = virtnet_probe, 3083 .remove = virtnet_remove, 3084 .config_changed = virtnet_config_changed, 3085 #ifdef CONFIG_PM_SLEEP 3086 .freeze = virtnet_freeze, 3087 .restore = virtnet_restore, 3088 #endif 3089 }; 3090 3091 static __init int virtio_net_driver_init(void) 3092 { 3093 int ret; 3094 3095 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 3096 virtnet_cpu_online, 3097 virtnet_cpu_down_prep); 3098 if (ret < 0) 3099 goto out; 3100 virtionet_online = ret; 3101 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 3102 NULL, virtnet_cpu_dead); 3103 if (ret) 3104 goto err_dead; 3105 3106 ret = register_virtio_driver(&virtio_net_driver); 3107 if (ret) 3108 goto err_virtio; 3109 return 0; 3110 err_virtio: 3111 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 3112 err_dead: 3113 cpuhp_remove_multi_state(virtionet_online); 3114 out: 3115 return ret; 3116 } 3117 module_init(virtio_net_driver_init); 3118 3119 static __exit void virtio_net_driver_exit(void) 3120 { 3121 unregister_virtio_driver(&virtio_net_driver); 3122 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 3123 cpuhp_remove_multi_state(virtionet_online); 3124 } 3125 module_exit(virtio_net_driver_exit); 3126 3127 MODULE_DEVICE_TABLE(virtio, id_table); 3128 MODULE_DESCRIPTION("Virtio network driver"); 3129 MODULE_LICENSE("GPL"); 3130