net.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/* Copyright (C) 2009 Red Hat, Inc.
   3 * Author: Michael S. Tsirkin <mst@redhat.com>
   4 *
   5 * virtio-net server in host kernel.
   6 */
   7
   8#include <linux/compat.h>
   9#include <linux/eventfd.h>
  10#include <linux/vhost.h>
  11#include <linux/virtio_net.h>
  12#include <linux/miscdevice.h>
  13#include <linux/module.h>
  14#include <linux/moduleparam.h>
  15#include <linux/mutex.h>
  16#include <linux/workqueue.h>
  17#include <linux/file.h>
  18#include <linux/slab.h>
  19#include <linux/sched/clock.h>
  20#include <linux/sched/signal.h>
  21#include <linux/vmalloc.h>
  22
  23#include <linux/net.h>
  24#include <linux/if_packet.h>
  25#include <linux/if_arp.h>
  26#include <linux/if_tun.h>
  27#include <linux/if_macvlan.h>
  28#include <linux/if_tap.h>
  29#include <linux/if_vlan.h>
  30#include <linux/skb_array.h>
  31#include <linux/skbuff.h>
  32
  33#include <net/sock.h>
  34#include <net/xdp.h>
  35
  36#include "vhost.h"
  37
  38static int experimental_zcopytx = 0;
  39module_param(experimental_zcopytx, int, 0444);
  40MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  41		                       " 1 -Enable; 0 - Disable");
  42
  43/* Max number of bytes transferred before requeueing the job.
  44 * Using this limit prevents one virtqueue from starving others. */
  45#define VHOST_NET_WEIGHT 0x80000
  46
  47/* Max number of packets transferred before requeueing the job.
  48 * Using this limit prevents one virtqueue from starving others with small
  49 * pkts.
  50 */
  51#define VHOST_NET_PKT_WEIGHT 256
  52
  53/* MAX number of TX used buffers for outstanding zerocopy */
  54#define VHOST_MAX_PEND 128
  55#define VHOST_GOODCOPY_LEN 256
  56
  57/*
  58 * For transmit, used buffer len is unused; we override it to track buffer
  59 * status internally; used for zerocopy tx only.
  60 */
  61/* Lower device DMA failed */
  62#define VHOST_DMA_FAILED_LEN	((__force __virtio32)3)
  63/* Lower device DMA done */
  64#define VHOST_DMA_DONE_LEN	((__force __virtio32)2)
  65/* Lower device DMA in progress */
  66#define VHOST_DMA_IN_PROGRESS	((__force __virtio32)1)
  67/* Buffer unused */
  68#define VHOST_DMA_CLEAR_LEN	((__force __virtio32)0)
  69
  70#define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  71
  72static const int vhost_net_bits[] = {
  73	VHOST_FEATURES,
  74	VHOST_NET_F_VIRTIO_NET_HDR,
  75	VIRTIO_NET_F_MRG_RXBUF,
  76	VIRTIO_F_ACCESS_PLATFORM,
  77	VIRTIO_F_RING_RESET,
  78	VIRTIO_F_IN_ORDER,
  79	VIRTIO_NET_F_GUEST_UDP_TUNNEL_GSO,
  80	VIRTIO_NET_F_HOST_UDP_TUNNEL_GSO
  81};
  82
  83enum {
  84	VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  85};
  86
  87enum {
  88	VHOST_NET_VQ_RX = 0,
  89	VHOST_NET_VQ_TX = 1,
  90	VHOST_NET_VQ_MAX = 2,
  91};
  92
  93struct vhost_net_ubuf_ref {
  94	/* refcount follows semantics similar to kref:
  95	 *  0: object is released
  96	 *  1: no outstanding ubufs
  97	 * >1: outstanding ubufs
  98	 */
  99	atomic_t refcount;
 100	wait_queue_head_t wait;
 101	struct vhost_virtqueue *vq;
 102	struct rcu_head rcu;
 103};
 104
 105#define VHOST_NET_BATCH 64
 106struct vhost_net_buf {
 107	void **queue;
 108	int tail;
 109	int head;
 110};
 111
 112struct vhost_net_virtqueue {
 113	struct vhost_virtqueue vq;
 114	size_t vhost_hlen;
 115	size_t sock_hlen;
 116	/* vhost zerocopy support fields below: */
 117	/* last used idx for outstanding DMA zerocopy buffers */
 118	int upend_idx;
 119	/* For TX, first used idx for DMA done zerocopy buffers
 120	 * For RX, number of batched heads
 121	 */
 122	int done_idx;
 123	/* Number of XDP frames batched */
 124	int batched_xdp;
 125	/* an array of userspace buffers info */
 126	struct ubuf_info_msgzc *ubuf_info;
 127	/* Reference counting for outstanding ubufs.
 128	 * Protected by vq mutex. Writers must also take device mutex. */
 129	struct vhost_net_ubuf_ref *ubufs;
 130	struct ptr_ring *rx_ring;
 131	struct vhost_net_buf rxq;
 132	/* Batched XDP buffs */
 133	struct xdp_buff *xdp;
 134};
 135
 136struct vhost_net {
 137	struct vhost_dev dev;
 138	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 139	struct vhost_poll poll[VHOST_NET_VQ_MAX];
 140	/* Number of TX recently submitted.
 141	 * Protected by tx vq lock. */
 142	unsigned tx_packets;
 143	/* Number of times zerocopy TX recently failed.
 144	 * Protected by tx vq lock. */
 145	unsigned tx_zcopy_err;
 146	/* Flush in progress. Protected by tx vq lock. */
 147	bool tx_flush;
 148	/* Private page frag cache */
 149	struct page_frag_cache pf_cache;
 150};
 151
 152static unsigned vhost_net_zcopy_mask __read_mostly;
 153
 154static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 155{
 156	if (rxq->tail != rxq->head)
 157		return rxq->queue[rxq->head];
 158	else
 159		return NULL;
 160}
 161
 162static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 163{
 164	return rxq->tail - rxq->head;
 165}
 166
 167static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 168{
 169	return rxq->tail == rxq->head;
 170}
 171
 172static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 173{
 174	void *ret = vhost_net_buf_get_ptr(rxq);
 175	++rxq->head;
 176	return ret;
 177}
 178
 179static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 180{
 181	struct vhost_net_buf *rxq = &nvq->rxq;
 182
 183	rxq->head = 0;
 184	rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 185					      VHOST_NET_BATCH);
 186	return rxq->tail;
 187}
 188
 189static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 190{
 191	struct vhost_net_buf *rxq = &nvq->rxq;
 192
 193	if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 194		ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 195				   vhost_net_buf_get_size(rxq),
 196				   tun_ptr_free);
 197		rxq->head = rxq->tail = 0;
 198	}
 199}
 200
 201static int vhost_net_buf_peek_len(void *ptr)
 202{
 203	if (tun_is_xdp_frame(ptr)) {
 204		struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 205
 206		return xdpf->len;
 207	}
 208
 209	return __skb_array_len_with_tag(ptr);
 210}
 211
 212static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 213{
 214	struct vhost_net_buf *rxq = &nvq->rxq;
 215
 216	if (!vhost_net_buf_is_empty(rxq))
 217		goto out;
 218
 219	if (!vhost_net_buf_produce(nvq))
 220		return 0;
 221
 222out:
 223	return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 224}
 225
 226static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 227{
 228	rxq->head = rxq->tail = 0;
 229}
 230
 231static void vhost_net_enable_zcopy(int vq)
 232{
 233	vhost_net_zcopy_mask |= 0x1 << vq;
 234}
 235
 236static struct vhost_net_ubuf_ref *
 237vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 238{
 239	struct vhost_net_ubuf_ref *ubufs;
 240	/* No zero copy backend? Nothing to count. */
 241	if (!zcopy)
 242		return NULL;
 243	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 244	if (!ubufs)
 245		return ERR_PTR(-ENOMEM);
 246	atomic_set(&ubufs->refcount, 1);
 247	init_waitqueue_head(&ubufs->wait);
 248	ubufs->vq = vq;
 249	return ubufs;
 250}
 251
 252static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 253{
 254	int r;
 255
 256	rcu_read_lock();
 257	r = atomic_sub_return(1, &ubufs->refcount);
 258	if (unlikely(!r))
 259		wake_up(&ubufs->wait);
 260	rcu_read_unlock();
 261	return r;
 262}
 263
 264static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 265{
 266	vhost_net_ubuf_put(ubufs);
 267	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 268}
 269
 270static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 271{
 272	vhost_net_ubuf_put_and_wait(ubufs);
 273	kfree_rcu(ubufs, rcu);
 274}
 275
 276static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 277{
 278	int i;
 279
 280	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 281		kfree(n->vqs[i].ubuf_info);
 282		n->vqs[i].ubuf_info = NULL;
 283	}
 284}
 285
 286static int vhost_net_set_ubuf_info(struct vhost_net *n)
 287{
 288	bool zcopy;
 289	int i;
 290
 291	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 292		zcopy = vhost_net_zcopy_mask & (0x1 << i);
 293		if (!zcopy)
 294			continue;
 295		n->vqs[i].ubuf_info =
 296			kmalloc_array(UIO_MAXIOV,
 297				      sizeof(*n->vqs[i].ubuf_info),
 298				      GFP_KERNEL);
 299		if  (!n->vqs[i].ubuf_info)
 300			goto err;
 301	}
 302	return 0;
 303
 304err:
 305	vhost_net_clear_ubuf_info(n);
 306	return -ENOMEM;
 307}
 308
 309static void vhost_net_vq_reset(struct vhost_net *n)
 310{
 311	int i;
 312
 313	vhost_net_clear_ubuf_info(n);
 314
 315	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 316		n->vqs[i].done_idx = 0;
 317		n->vqs[i].upend_idx = 0;
 318		n->vqs[i].ubufs = NULL;
 319		n->vqs[i].vhost_hlen = 0;
 320		n->vqs[i].sock_hlen = 0;
 321		vhost_net_buf_init(&n->vqs[i].rxq);
 322	}
 323
 324}
 325
 326static void vhost_net_tx_packet(struct vhost_net *net)
 327{
 328	++net->tx_packets;
 329	if (net->tx_packets < 1024)
 330		return;
 331	net->tx_packets = 0;
 332	net->tx_zcopy_err = 0;
 333}
 334
 335static void vhost_net_tx_err(struct vhost_net *net)
 336{
 337	++net->tx_zcopy_err;
 338}
 339
 340static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 341{
 342	/* TX flush waits for outstanding DMAs to be done.
 343	 * Don't start new DMAs.
 344	 */
 345	return !net->tx_flush &&
 346		net->tx_packets / 64 >= net->tx_zcopy_err;
 347}
 348
 349static bool vhost_sock_zcopy(struct socket *sock)
 350{
 351	return unlikely(experimental_zcopytx) &&
 352		sock_flag(sock->sk, SOCK_ZEROCOPY);
 353}
 354
 355static bool vhost_sock_xdp(struct socket *sock)
 356{
 357	return sock_flag(sock->sk, SOCK_XDP);
 358}
 359
 360/* In case of DMA done not in order in lower device driver for some reason.
 361 * upend_idx is used to track end of used idx, done_idx is used to track head
 362 * of used idx. Once lower device DMA done contiguously, we will signal KVM
 363 * guest used idx.
 364 */
 365static void vhost_zerocopy_signal_used(struct vhost_net *net,
 366				       struct vhost_virtqueue *vq)
 367{
 368	struct vhost_net_virtqueue *nvq =
 369		container_of(vq, struct vhost_net_virtqueue, vq);
 370	int i, add;
 371	int j = 0;
 372
 373	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 374		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 375			vhost_net_tx_err(net);
 376		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 377			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 378			++j;
 379		} else
 380			break;
 381	}
 382	while (j) {
 383		add = min(UIO_MAXIOV - nvq->done_idx, j);
 384		vhost_add_used_and_signal_n(vq->dev, vq,
 385					    &vq->heads[nvq->done_idx],
 386					    NULL, add);
 387		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 388		j -= add;
 389	}
 390}
 391
 392static void vhost_zerocopy_complete(struct sk_buff *skb,
 393				    struct ubuf_info *ubuf_base, bool success)
 394{
 395	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
 396	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 397	struct vhost_virtqueue *vq = ubufs->vq;
 398	int cnt;
 399
 400	rcu_read_lock_bh();
 401
 402	/* set len to mark this desc buffers done DMA */
 403	vq->heads[ubuf->desc].len = success ?
 404		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 405	cnt = vhost_net_ubuf_put(ubufs);
 406
 407	/*
 408	 * Trigger polling thread if guest stopped submitting new buffers:
 409	 * in this case, the refcount after decrement will eventually reach 1.
 410	 * We also trigger polling periodically after each 16 packets
 411	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
 412	 * less than 10% of times).
 413	 */
 414	if (cnt <= 1 || !(cnt % 16))
 415		vhost_poll_queue(&vq->poll);
 416
 417	rcu_read_unlock_bh();
 418}
 419
 420static const struct ubuf_info_ops vhost_ubuf_ops = {
 421	.complete = vhost_zerocopy_complete,
 422};
 423
 424static inline unsigned long busy_clock(void)
 425{
 426	return local_clock() >> 10;
 427}
 428
 429static bool vhost_can_busy_poll(unsigned long endtime)
 430{
 431	return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 432		      !signal_pending(current));
 433}
 434
 435static void vhost_net_disable_vq(struct vhost_net *n,
 436				 struct vhost_virtqueue *vq)
 437{
 438	struct vhost_net_virtqueue *nvq =
 439		container_of(vq, struct vhost_net_virtqueue, vq);
 440	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 441	if (!vhost_vq_get_backend(vq))
 442		return;
 443	vhost_poll_stop(poll);
 444}
 445
 446static int vhost_net_enable_vq(struct vhost_net *n,
 447				struct vhost_virtqueue *vq)
 448{
 449	struct vhost_net_virtqueue *nvq =
 450		container_of(vq, struct vhost_net_virtqueue, vq);
 451	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 452	struct socket *sock;
 453
 454	sock = vhost_vq_get_backend(vq);
 455	if (!sock)
 456		return 0;
 457
 458	return vhost_poll_start(poll, sock->file);
 459}
 460
 461static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq,
 462				  unsigned int count)
 463{
 464	struct vhost_virtqueue *vq = &nvq->vq;
 465	struct vhost_dev *dev = vq->dev;
 466
 467	if (!nvq->done_idx)
 468		return;
 469
 470	vhost_add_used_and_signal_n(dev, vq, vq->heads,
 471				    vq->nheads, count);
 472	nvq->done_idx = 0;
 473}
 474
 475static void vhost_tx_batch(struct vhost_net *net,
 476			   struct vhost_net_virtqueue *nvq,
 477			   struct socket *sock,
 478			   struct msghdr *msghdr)
 479{
 480	struct vhost_virtqueue *vq = &nvq->vq;
 481	bool in_order = vhost_has_feature(vq, VIRTIO_F_IN_ORDER);
 482	struct tun_msg_ctl ctl = {
 483		.type = TUN_MSG_PTR,
 484		.num = nvq->batched_xdp,
 485		.ptr = nvq->xdp,
 486	};
 487	int i, err;
 488
 489	if (in_order) {
 490		vq->heads[0].len = 0;
 491		vq->nheads[0] = nvq->done_idx;
 492	}
 493
 494	if (nvq->batched_xdp == 0)
 495		goto signal_used;
 496
 497	msghdr->msg_control = &ctl;
 498	msghdr->msg_controllen = sizeof(ctl);
 499	err = sock->ops->sendmsg(sock, msghdr, 0);
 500	if (unlikely(err < 0)) {
 501		vq_err(&nvq->vq, "Fail to batch sending packets\n");
 502
 503		/* free pages owned by XDP; since this is an unlikely error path,
 504		 * keep it simple and avoid more complex bulk update for the
 505		 * used pages
 506		 */
 507		for (i = 0; i < nvq->batched_xdp; ++i)
 508			put_page(virt_to_head_page(nvq->xdp[i].data));
 509		nvq->batched_xdp = 0;
 510		nvq->done_idx = 0;
 511		return;
 512	}
 513
 514signal_used:
 515	vhost_net_signal_used(nvq, in_order ? 1 : nvq->done_idx);
 516	nvq->batched_xdp = 0;
 517}
 518
 519static int sock_has_rx_data(struct socket *sock)
 520{
 521	if (unlikely(!sock))
 522		return 0;
 523
 524	if (sock->ops->peek_len)
 525		return sock->ops->peek_len(sock);
 526
 527	return skb_queue_empty(&sock->sk->sk_receive_queue);
 528}
 529
 530static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 531					  struct vhost_virtqueue *vq)
 532{
 533	if (!vhost_vq_avail_empty(&net->dev, vq)) {
 534		vhost_poll_queue(&vq->poll);
 535	} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 536		vhost_disable_notify(&net->dev, vq);
 537		vhost_poll_queue(&vq->poll);
 538	}
 539}
 540
 541static void vhost_net_busy_poll(struct vhost_net *net,
 542				struct vhost_virtqueue *rvq,
 543				struct vhost_virtqueue *tvq,
 544				bool *busyloop_intr,
 545				bool poll_rx)
 546{
 547	unsigned long busyloop_timeout;
 548	unsigned long endtime;
 549	struct socket *sock;
 550	struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 551
 552	/* Try to hold the vq mutex of the paired virtqueue. We can't
 553	 * use mutex_lock() here since we could not guarantee a
 554	 * consistenet lock ordering.
 555	 */
 556	if (!mutex_trylock(&vq->mutex))
 557		return;
 558
 559	vhost_disable_notify(&net->dev, vq);
 560	sock = vhost_vq_get_backend(rvq);
 561
 562	busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 563				     tvq->busyloop_timeout;
 564
 565	preempt_disable();
 566	endtime = busy_clock() + busyloop_timeout;
 567
 568	while (vhost_can_busy_poll(endtime)) {
 569		if (vhost_vq_has_work(vq)) {
 570			*busyloop_intr = true;
 571			break;
 572		}
 573
 574		if ((sock_has_rx_data(sock) &&
 575		     !vhost_vq_avail_empty(&net->dev, rvq)) ||
 576		    !vhost_vq_avail_empty(&net->dev, tvq))
 577			break;
 578
 579		cpu_relax();
 580	}
 581
 582	preempt_enable();
 583
 584	if (poll_rx || sock_has_rx_data(sock))
 585		vhost_net_busy_poll_try_queue(net, vq);
 586	else if (!poll_rx) /* On tx here, sock has no rx data. */
 587		vhost_enable_notify(&net->dev, rvq);
 588
 589	mutex_unlock(&vq->mutex);
 590}
 591
 592static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 593				    struct vhost_net_virtqueue *tnvq,
 594				    unsigned int *out_num, unsigned int *in_num,
 595				    struct msghdr *msghdr, bool *busyloop_intr,
 596				    unsigned int *ndesc)
 597{
 598	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 599	struct vhost_virtqueue *rvq = &rnvq->vq;
 600	struct vhost_virtqueue *tvq = &tnvq->vq;
 601
 602	int r = vhost_get_vq_desc_n(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 603				    out_num, in_num, NULL, NULL, ndesc);
 604
 605	if (r == tvq->num && tvq->busyloop_timeout) {
 606		/* Flush batched packets first */
 607		if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
 608			vhost_tx_batch(net, tnvq,
 609				       vhost_vq_get_backend(tvq),
 610				       msghdr);
 611
 612		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 613
 614		r = vhost_get_vq_desc_n(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 615					out_num, in_num, NULL, NULL, ndesc);
 616	}
 617
 618	return r;
 619}
 620
 621static bool vhost_exceeds_maxpend(struct vhost_net *net)
 622{
 623	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 624	struct vhost_virtqueue *vq = &nvq->vq;
 625
 626	return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 627	       min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 628}
 629
 630static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 631			    size_t hdr_size, int out)
 632{
 633	/* Skip header. TODO: support TSO. */
 634	size_t len = iov_length(vq->iov, out);
 635
 636	iov_iter_init(iter, ITER_SOURCE, vq->iov, out, len);
 637	iov_iter_advance(iter, hdr_size);
 638
 639	return iov_iter_count(iter);
 640}
 641
 642static int get_tx_bufs(struct vhost_net *net,
 643		       struct vhost_net_virtqueue *nvq,
 644		       struct msghdr *msg,
 645		       unsigned int *out, unsigned int *in,
 646		       size_t *len, bool *busyloop_intr,
 647		       unsigned int *ndesc)
 648{
 649	struct vhost_virtqueue *vq = &nvq->vq;
 650	int ret;
 651
 652	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg,
 653				       busyloop_intr, ndesc);
 654
 655	if (ret < 0 || ret == vq->num)
 656		return ret;
 657
 658	if (*in) {
 659		vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 660			*out, *in);
 661		return -EFAULT;
 662	}
 663
 664	/* Sanity check */
 665	*len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 666	if (*len == 0) {
 667		vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 668			*len, nvq->vhost_hlen);
 669		return -EFAULT;
 670	}
 671
 672	return ret;
 673}
 674
 675static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 676{
 677	return total_len < VHOST_NET_WEIGHT &&
 678	       !vhost_vq_avail_empty(vq->dev, vq);
 679}
 680
 681#define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 682
 683static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 684			       struct iov_iter *from)
 685{
 686	struct vhost_virtqueue *vq = &nvq->vq;
 687	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 688					     dev);
 689	struct socket *sock = vhost_vq_get_backend(vq);
 690	struct virtio_net_hdr *gso;
 691	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 692	size_t len = iov_iter_count(from);
 693	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 694	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 695	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 696	int sock_hlen = nvq->sock_hlen;
 697	void *buf;
 698	int copied;
 699	int ret;
 700
 701	if (unlikely(len < nvq->sock_hlen))
 702		return -EFAULT;
 703
 704	if (SKB_DATA_ALIGN(len + pad) +
 705	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 706		return -ENOSPC;
 707
 708	buflen += SKB_DATA_ALIGN(len + pad);
 709	buf = page_frag_alloc_align(&net->pf_cache, buflen, GFP_KERNEL,
 710				    SMP_CACHE_BYTES);
 711	if (unlikely(!buf))
 712		return -ENOMEM;
 713
 714	copied = copy_from_iter(buf + pad - sock_hlen, len, from);
 715	if (copied != len) {
 716		ret = -EFAULT;
 717		goto err;
 718	}
 719
 720	gso = buf + pad - sock_hlen;
 721
 722	if (!sock_hlen)
 723		memset(buf, 0, pad);
 724
 725	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 726	    vhost16_to_cpu(vq, gso->csum_start) +
 727	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 728	    vhost16_to_cpu(vq, gso->hdr_len)) {
 729		gso->hdr_len = cpu_to_vhost16(vq,
 730			       vhost16_to_cpu(vq, gso->csum_start) +
 731			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
 732
 733		if (vhost16_to_cpu(vq, gso->hdr_len) > len) {
 734			ret = -EINVAL;
 735			goto err;
 736		}
 737	}
 738
 739	/* pad contains sock_hlen */
 740	memcpy(buf, buf + pad - sock_hlen, sock_hlen);
 741
 742	xdp_init_buff(xdp, buflen, NULL);
 743	xdp_prepare_buff(xdp, buf, pad, len - sock_hlen, true);
 744
 745	++nvq->batched_xdp;
 746
 747	return 0;
 748
 749err:
 750	page_frag_free(buf);
 751	return ret;
 752}
 753
 754static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 755{
 756	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 757	struct vhost_virtqueue *vq = &nvq->vq;
 758	unsigned out, in;
 759	int head;
 760	struct msghdr msg = {
 761		.msg_name = NULL,
 762		.msg_namelen = 0,
 763		.msg_control = NULL,
 764		.msg_controllen = 0,
 765		.msg_flags = MSG_DONTWAIT,
 766	};
 767	size_t len, total_len = 0;
 768	int err;
 769	int sent_pkts = 0;
 770	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 771	bool in_order = vhost_has_feature(vq, VIRTIO_F_IN_ORDER);
 772	unsigned int ndesc = 0;
 773
 774	do {
 775		bool busyloop_intr = false;
 776
 777		if (nvq->done_idx == VHOST_NET_BATCH)
 778			vhost_tx_batch(net, nvq, sock, &msg);
 779
 780		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 781				   &busyloop_intr, &ndesc);
 782		/* On error, stop handling until the next kick. */
 783		if (unlikely(head < 0))
 784			break;
 785		/* Nothing new?  Wait for eventfd to tell us they refilled. */
 786		if (head == vq->num) {
 787			/* Flush batched packets to handle pending RX
 788			 * work (if busyloop_intr is set) and to avoid
 789			 * unnecessary virtqueue kicks.
 790			 */
 791			vhost_tx_batch(net, nvq, sock, &msg);
 792			if (unlikely(busyloop_intr)) {
 793				vhost_poll_queue(&vq->poll);
 794			} else if (unlikely(vhost_enable_notify(&net->dev,
 795								vq))) {
 796				vhost_disable_notify(&net->dev, vq);
 797				continue;
 798			}
 799			break;
 800		}
 801
 802		total_len += len;
 803
 804		/* For simplicity, TX batching is only enabled if
 805		 * sndbuf is unlimited.
 806		 */
 807		if (sock_can_batch) {
 808			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 809			if (!err) {
 810				goto done;
 811			} else if (unlikely(err != -ENOSPC)) {
 812				vhost_tx_batch(net, nvq, sock, &msg);
 813				vhost_discard_vq_desc(vq, 1, ndesc);
 814				vhost_net_enable_vq(net, vq);
 815				break;
 816			}
 817
 818			if (nvq->batched_xdp) {
 819				/* We can't build XDP buff, go for single
 820				 * packet path but let's flush batched
 821				 * packets.
 822				 */
 823				vhost_tx_batch(net, nvq, sock, &msg);
 824			}
 825			msg.msg_control = NULL;
 826		} else {
 827			if (tx_can_batch(vq, total_len))
 828				msg.msg_flags |= MSG_MORE;
 829			else
 830				msg.msg_flags &= ~MSG_MORE;
 831		}
 832
 833		err = sock->ops->sendmsg(sock, &msg, len);
 834		if (unlikely(err < 0)) {
 835			if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
 836				vhost_discard_vq_desc(vq, 1, ndesc);
 837				vhost_net_enable_vq(net, vq);
 838				break;
 839			}
 840			pr_debug("Fail to send packet: err %d", err);
 841		} else if (unlikely(err != len))
 842			pr_debug("Truncated TX packet: len %d != %zd\n",
 843				 err, len);
 844done:
 845		if (in_order) {
 846			vq->heads[0].id = cpu_to_vhost32(vq, head);
 847		} else {
 848			vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 849			vq->heads[nvq->done_idx].len = 0;
 850		}
 851		++nvq->done_idx;
 852	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 853
 854	vhost_tx_batch(net, nvq, sock, &msg);
 855}
 856
 857static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 858{
 859	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 860	struct vhost_virtqueue *vq = &nvq->vq;
 861	unsigned out, in;
 862	int head;
 863	struct msghdr msg = {
 864		.msg_name = NULL,
 865		.msg_namelen = 0,
 866		.msg_control = NULL,
 867		.msg_controllen = 0,
 868		.msg_flags = MSG_DONTWAIT,
 869	};
 870	struct tun_msg_ctl ctl;
 871	size_t len, total_len = 0;
 872	int err;
 873	struct vhost_net_ubuf_ref *ubufs;
 874	struct ubuf_info_msgzc *ubuf;
 875	unsigned int ndesc = 0;
 876	bool zcopy_used;
 877	int sent_pkts = 0;
 878
 879	do {
 880		bool busyloop_intr;
 881
 882		/* Release DMAs done buffers first */
 883		vhost_zerocopy_signal_used(net, vq);
 884
 885		busyloop_intr = false;
 886		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 887				   &busyloop_intr, &ndesc);
 888		/* On error, stop handling until the next kick. */
 889		if (unlikely(head < 0))
 890			break;
 891		/* Nothing new?  Wait for eventfd to tell us they refilled. */
 892		if (head == vq->num) {
 893			if (unlikely(busyloop_intr)) {
 894				vhost_poll_queue(&vq->poll);
 895			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 896				vhost_disable_notify(&net->dev, vq);
 897				continue;
 898			}
 899			break;
 900		}
 901
 902		zcopy_used = len >= VHOST_GOODCOPY_LEN
 903			     && !vhost_exceeds_maxpend(net)
 904			     && vhost_net_tx_select_zcopy(net);
 905
 906		/* use msg_control to pass vhost zerocopy ubuf info to skb */
 907		if (zcopy_used) {
 908			ubuf = nvq->ubuf_info + nvq->upend_idx;
 909			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 910			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 911			ubuf->ctx = nvq->ubufs;
 912			ubuf->desc = nvq->upend_idx;
 913			ubuf->ubuf.ops = &vhost_ubuf_ops;
 914			ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
 915			refcount_set(&ubuf->ubuf.refcnt, 1);
 916			msg.msg_control = &ctl;
 917			ctl.type = TUN_MSG_UBUF;
 918			ctl.ptr = &ubuf->ubuf;
 919			msg.msg_controllen = sizeof(ctl);
 920			ubufs = nvq->ubufs;
 921			atomic_inc(&ubufs->refcount);
 922			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 923		} else {
 924			msg.msg_control = NULL;
 925			ubufs = NULL;
 926		}
 927		total_len += len;
 928		if (tx_can_batch(vq, total_len) &&
 929		    likely(!vhost_exceeds_maxpend(net))) {
 930			msg.msg_flags |= MSG_MORE;
 931		} else {
 932			msg.msg_flags &= ~MSG_MORE;
 933		}
 934
 935		err = sock->ops->sendmsg(sock, &msg, len);
 936		if (unlikely(err < 0)) {
 937			bool retry = err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS;
 938
 939			if (zcopy_used) {
 940				if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 941					vhost_net_ubuf_put(ubufs);
 942				if (retry)
 943					nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 944						% UIO_MAXIOV;
 945				else
 946					vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
 947			}
 948			if (retry) {
 949				vhost_discard_vq_desc(vq, 1, ndesc);
 950				vhost_net_enable_vq(net, vq);
 951				break;
 952			}
 953			pr_debug("Fail to send packet: err %d", err);
 954		} else if (unlikely(err != len))
 955			pr_debug("Truncated TX packet: "
 956				 " len %d != %zd\n", err, len);
 957		if (!zcopy_used)
 958			vhost_add_used_and_signal(&net->dev, vq, head, 0);
 959		else
 960			vhost_zerocopy_signal_used(net, vq);
 961		vhost_net_tx_packet(net);
 962	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 963}
 964
 965/* Expects to be always run from workqueue - which acts as
 966 * read-size critical section for our kind of RCU. */
 967static void handle_tx(struct vhost_net *net)
 968{
 969	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 970	struct vhost_virtqueue *vq = &nvq->vq;
 971	struct socket *sock;
 972
 973	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 974	sock = vhost_vq_get_backend(vq);
 975	if (!sock)
 976		goto out;
 977
 978	if (!vq_meta_prefetch(vq))
 979		goto out;
 980
 981	vhost_disable_notify(&net->dev, vq);
 982	vhost_net_disable_vq(net, vq);
 983
 984	if (vhost_sock_zcopy(sock))
 985		handle_tx_zerocopy(net, sock);
 986	else
 987		handle_tx_copy(net, sock);
 988
 989out:
 990	mutex_unlock(&vq->mutex);
 991}
 992
 993static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 994{
 995	struct sk_buff *head;
 996	int len = 0;
 997	unsigned long flags;
 998
 999	if (rvq->rx_ring)
1000		return vhost_net_buf_peek(rvq);
1001
1002	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
1003	head = skb_peek(&sk->sk_receive_queue);
1004	if (likely(head)) {
1005		len = head->len;
1006		if (skb_vlan_tag_present(head))
1007			len += VLAN_HLEN;
1008	}
1009
1010	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1011	return len;
1012}
1013
1014static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1015				      bool *busyloop_intr, unsigned int *count)
1016{
1017	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1018	struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1019	struct vhost_virtqueue *rvq = &rnvq->vq;
1020	struct vhost_virtqueue *tvq = &tnvq->vq;
1021	int len = peek_head_len(rnvq, sk);
1022
1023	if (!len && rvq->busyloop_timeout) {
1024		/* Flush batched heads first */
1025		vhost_net_signal_used(rnvq, *count);
1026		*count = 0;
1027		/* Both tx vq and rx socket were polled here */
1028		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1029
1030		len = peek_head_len(rnvq, sk);
1031	}
1032
1033	return len;
1034}
1035
1036/* This is a multi-buffer version of vhost_get_desc, that works if
1037 *	vq has read descriptors only.
1038 * @nvq		- the relevant vhost_net virtqueue
1039 * @datalen	- data length we'll be reading
1040 * @iovcount	- returned count of io vectors we fill
1041 * @log		- vhost log
1042 * @log_num	- log offset
1043 * @quota       - headcount quota, 1 for big buffer
1044 *	returns number of buffer heads allocated, negative on error
1045 */
1046static int get_rx_bufs(struct vhost_net_virtqueue *nvq,
1047		       struct vring_used_elem *heads,
1048		       u16 *nheads,
1049		       int datalen,
1050		       unsigned *iovcount,
1051		       struct vhost_log *log,
1052		       unsigned *log_num,
1053		       unsigned int quota,
1054		       unsigned int *ndesc)
1055{
1056	struct vhost_virtqueue *vq = &nvq->vq;
1057	bool in_order = vhost_has_feature(vq, VIRTIO_F_IN_ORDER);
1058	unsigned int out, in, desc_num, n = 0;
1059	int seg = 0;
1060	int headcount = 0;
1061	unsigned d;
1062	int r, nlogs = 0;
1063	/* len is always initialized before use since we are always called with
1064	 * datalen > 0.
1065	 */
1066	u32 len;
1067
1068	while (datalen > 0 && headcount < quota) {
1069		if (unlikely(seg >= UIO_MAXIOV)) {
1070			r = -ENOBUFS;
1071			goto err;
1072		}
1073		r = vhost_get_vq_desc_n(vq, vq->iov + seg,
1074					ARRAY_SIZE(vq->iov) - seg, &out,
1075					&in, log, log_num, &desc_num);
1076		if (unlikely(r < 0))
1077			goto err;
1078
1079		d = r;
1080		if (d == vq->num) {
1081			r = 0;
1082			goto err;
1083		}
1084		if (unlikely(out || in <= 0)) {
1085			vq_err(vq, "unexpected descriptor format for RX: "
1086				"out %d, in %d\n", out, in);
1087			r = -EINVAL;
1088			goto err;
1089		}
1090		if (unlikely(log)) {
1091			nlogs += *log_num;
1092			log += *log_num;
1093		}
1094		len = iov_length(vq->iov + seg, in);
1095		if (!in_order) {
1096			heads[headcount].id = cpu_to_vhost32(vq, d);
1097			heads[headcount].len = cpu_to_vhost32(vq, len);
1098		}
1099		++headcount;
1100		datalen -= len;
1101		seg += in;
1102		n += desc_num;
1103	}
1104
1105	*iovcount = seg;
1106	if (unlikely(log))
1107		*log_num = nlogs;
1108
1109	/* Detect overrun */
1110	if (unlikely(datalen > 0)) {
1111		r = UIO_MAXIOV + 1;
1112		goto err;
1113	}
1114
1115	if (!in_order)
1116		heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1117	else {
1118		heads[0].len = cpu_to_vhost32(vq, len + datalen);
1119		heads[0].id = cpu_to_vhost32(vq, d);
1120		nheads[0] = headcount;
1121	}
1122
1123	*ndesc = n;
1124
1125	return headcount;
1126err:
1127	vhost_discard_vq_desc(vq, headcount, n);
1128	return r;
1129}
1130
1131/* Expects to be always run from workqueue - which acts as
1132 * read-size critical section for our kind of RCU. */
1133static void handle_rx(struct vhost_net *net)
1134{
1135	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1136	struct vhost_virtqueue *vq = &nvq->vq;
1137	bool in_order = vhost_has_feature(vq, VIRTIO_F_IN_ORDER);
1138	unsigned int count = 0;
1139	unsigned in, log;
1140	struct vhost_log *vq_log;
1141	struct msghdr msg = {
1142		.msg_name = NULL,
1143		.msg_namelen = 0,
1144		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
1145		.msg_controllen = 0,
1146		.msg_flags = MSG_DONTWAIT,
1147	};
1148	struct virtio_net_hdr hdr = {
1149		.flags = 0,
1150		.gso_type = VIRTIO_NET_HDR_GSO_NONE
1151	};
1152	size_t total_len = 0;
1153	int err, mergeable;
1154	s16 headcount;
1155	size_t vhost_hlen, sock_hlen;
1156	size_t vhost_len, sock_len;
1157	bool busyloop_intr = false;
1158	bool set_num_buffers;
1159	struct socket *sock;
1160	struct iov_iter fixup;
1161	__virtio16 num_buffers;
1162	int recv_pkts = 0;
1163	unsigned int ndesc;
1164
1165	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1166	sock = vhost_vq_get_backend(vq);
1167	if (!sock)
1168		goto out;
1169
1170	if (!vq_meta_prefetch(vq))
1171		goto out;
1172
1173	vhost_disable_notify(&net->dev, vq);
1174	vhost_net_disable_vq(net, vq);
1175
1176	vhost_hlen = nvq->vhost_hlen;
1177	sock_hlen = nvq->sock_hlen;
1178
1179	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1180		vq->log : NULL;
1181	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1182	set_num_buffers = mergeable ||
1183			  vhost_has_feature(vq, VIRTIO_F_VERSION_1);
1184
1185	do {
1186		sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1187						      &busyloop_intr, &count);
1188		if (!sock_len)
1189			break;
1190		sock_len += sock_hlen;
1191		vhost_len = sock_len + vhost_hlen;
1192		headcount = get_rx_bufs(nvq, vq->heads + count,
1193					vq->nheads + count,
1194					vhost_len, &in, vq_log, &log,
1195					likely(mergeable) ? UIO_MAXIOV : 1,
1196					&ndesc);
1197		/* On error, stop handling until the next kick. */
1198		if (unlikely(headcount < 0))
1199			goto out;
1200		/* OK, now we need to know about added descriptors. */
1201		if (!headcount) {
1202			if (unlikely(busyloop_intr)) {
1203				vhost_poll_queue(&vq->poll);
1204			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1205				/* They have slipped one in as we were
1206				 * doing that: check again. */
1207				vhost_disable_notify(&net->dev, vq);
1208				continue;
1209			}
1210			/* Nothing new?  Wait for eventfd to tell us
1211			 * they refilled. */
1212			goto out;
1213		}
1214		busyloop_intr = false;
1215		if (nvq->rx_ring)
1216			msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1217		/* On overrun, truncate and discard */
1218		if (unlikely(headcount > UIO_MAXIOV)) {
1219			iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, 1, 1);
1220			err = sock->ops->recvmsg(sock, &msg,
1221						 1, MSG_DONTWAIT | MSG_TRUNC);
1222			pr_debug("Discarded rx packet: len %zd\n", sock_len);
1223			continue;
1224		}
1225		/* We don't need to be notified again. */
1226		iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, in, vhost_len);
1227		fixup = msg.msg_iter;
1228		if (unlikely((vhost_hlen))) {
1229			/* We will supply the header ourselves
1230			 * TODO: support TSO.
1231			 */
1232			iov_iter_advance(&msg.msg_iter, vhost_hlen);
1233		}
1234		err = sock->ops->recvmsg(sock, &msg,
1235					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1236		/* Userspace might have consumed the packet meanwhile:
1237		 * it's not supposed to do this usually, but might be hard
1238		 * to prevent. Discard data we got (if any) and keep going. */
1239		if (unlikely(err != sock_len)) {
1240			pr_debug("Discarded rx packet: "
1241				 " len %d, expected %zd\n", err, sock_len);
1242			vhost_discard_vq_desc(vq, headcount, ndesc);
1243			continue;
1244		}
1245		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1246		if (unlikely(vhost_hlen)) {
1247			if (copy_to_iter(&hdr, sizeof(hdr),
1248					 &fixup) != sizeof(hdr)) {
1249				vq_err(vq, "Unable to write vnet_hdr "
1250				       "at addr %p\n", vq->iov->iov_base);
1251				goto out;
1252			}
1253		} else {
1254			/* Header came from socket; we'll need to patch
1255			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1256			 */
1257			iov_iter_advance(&fixup, sizeof(hdr));
1258		}
1259		/* TODO: Should check and handle checksum. */
1260
1261		num_buffers = cpu_to_vhost16(vq, headcount);
1262		if (likely(set_num_buffers) &&
1263		    copy_to_iter(&num_buffers, sizeof num_buffers,
1264				 &fixup) != sizeof num_buffers) {
1265			vq_err(vq, "Failed num_buffers write");
1266			vhost_discard_vq_desc(vq, headcount, ndesc);
1267			goto out;
1268		}
1269		nvq->done_idx += headcount;
1270		count += in_order ? 1 : headcount;
1271		if (nvq->done_idx > VHOST_NET_BATCH) {
1272			vhost_net_signal_used(nvq, count);
1273			count = 0;
1274		}
1275		if (unlikely(vq_log))
1276			vhost_log_write(vq, vq_log, log, vhost_len,
1277					vq->iov, in);
1278		total_len += vhost_len;
1279	} while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1280
1281	if (unlikely(busyloop_intr))
1282		vhost_poll_queue(&vq->poll);
1283	else if (!sock_len)
1284		vhost_net_enable_vq(net, vq);
1285out:
1286	vhost_net_signal_used(nvq, count);
1287	mutex_unlock(&vq->mutex);
1288}
1289
1290static void handle_tx_kick(struct vhost_work *work)
1291{
1292	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1293						  poll.work);
1294	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1295
1296	handle_tx(net);
1297}
1298
1299static void handle_rx_kick(struct vhost_work *work)
1300{
1301	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1302						  poll.work);
1303	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1304
1305	handle_rx(net);
1306}
1307
1308static void handle_tx_net(struct vhost_work *work)
1309{
1310	struct vhost_net *net = container_of(work, struct vhost_net,
1311					     poll[VHOST_NET_VQ_TX].work);
1312	handle_tx(net);
1313}
1314
1315static void handle_rx_net(struct vhost_work *work)
1316{
1317	struct vhost_net *net = container_of(work, struct vhost_net,
1318					     poll[VHOST_NET_VQ_RX].work);
1319	handle_rx(net);
1320}
1321
1322static int vhost_net_open(struct inode *inode, struct file *f)
1323{
1324	struct vhost_net *n;
1325	struct vhost_dev *dev;
1326	struct vhost_virtqueue **vqs;
1327	void **queue;
1328	struct xdp_buff *xdp;
1329	int i;
1330
1331	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1332	if (!n)
1333		return -ENOMEM;
1334	vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1335	if (!vqs) {
1336		kvfree(n);
1337		return -ENOMEM;
1338	}
1339
1340	queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1341			      GFP_KERNEL);
1342	if (!queue) {
1343		kfree(vqs);
1344		kvfree(n);
1345		return -ENOMEM;
1346	}
1347	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1348
1349	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1350	if (!xdp) {
1351		kfree(vqs);
1352		kvfree(n);
1353		kfree(queue);
1354		return -ENOMEM;
1355	}
1356	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1357
1358	dev = &n->dev;
1359	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1360	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1361	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1362	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1363	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1364		n->vqs[i].ubufs = NULL;
1365		n->vqs[i].ubuf_info = NULL;
1366		n->vqs[i].upend_idx = 0;
1367		n->vqs[i].done_idx = 0;
1368		n->vqs[i].batched_xdp = 0;
1369		n->vqs[i].vhost_hlen = 0;
1370		n->vqs[i].sock_hlen = 0;
1371		n->vqs[i].rx_ring = NULL;
1372		vhost_net_buf_init(&n->vqs[i].rxq);
1373	}
1374	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1375		       UIO_MAXIOV + VHOST_NET_BATCH,
1376		       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1377		       NULL);
1378
1379	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev,
1380			vqs[VHOST_NET_VQ_TX]);
1381	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev,
1382			vqs[VHOST_NET_VQ_RX]);
1383
1384	f->private_data = n;
1385	page_frag_cache_init(&n->pf_cache);
1386
1387	return 0;
1388}
1389
1390static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1391					struct vhost_virtqueue *vq)
1392{
1393	struct socket *sock;
1394	struct vhost_net_virtqueue *nvq =
1395		container_of(vq, struct vhost_net_virtqueue, vq);
1396
1397	mutex_lock(&vq->mutex);
1398	sock = vhost_vq_get_backend(vq);
1399	vhost_net_disable_vq(n, vq);
1400	vhost_vq_set_backend(vq, NULL);
1401	vhost_net_buf_unproduce(nvq);
1402	nvq->rx_ring = NULL;
1403	mutex_unlock(&vq->mutex);
1404	return sock;
1405}
1406
1407static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1408			   struct socket **rx_sock)
1409{
1410	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1411	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1412}
1413
1414static void vhost_net_flush(struct vhost_net *n)
1415{
1416	vhost_dev_flush(&n->dev);
1417	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1418		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1419		n->tx_flush = true;
1420		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1421		/* Wait for all lower device DMAs done. */
1422		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1423		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1424		n->tx_flush = false;
1425		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1426		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1427	}
1428}
1429
1430static int vhost_net_release(struct inode *inode, struct file *f)
1431{
1432	struct vhost_net *n = f->private_data;
1433	struct socket *tx_sock;
1434	struct socket *rx_sock;
1435
1436	vhost_net_stop(n, &tx_sock, &rx_sock);
1437	vhost_net_flush(n);
1438	vhost_dev_stop(&n->dev);
1439	vhost_dev_cleanup(&n->dev);
1440	vhost_net_vq_reset(n);
1441	if (tx_sock)
1442		sockfd_put(tx_sock);
1443	if (rx_sock)
1444		sockfd_put(rx_sock);
1445	/* Make sure no callbacks are outstanding */
1446	synchronize_rcu();
1447	/* We do an extra flush before freeing memory,
1448	 * since jobs can re-queue themselves. */
1449	vhost_net_flush(n);
1450	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1451	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1452	kfree(n->dev.vqs);
1453	page_frag_cache_drain(&n->pf_cache);
1454	kvfree(n);
1455	return 0;
1456}
1457
1458static struct socket *get_raw_socket(int fd)
1459{
1460	int r;
1461	struct socket *sock = sockfd_lookup(fd, &r);
1462
1463	if (!sock)
1464		return ERR_PTR(-ENOTSOCK);
1465
1466	/* Parameter checking */
1467	if (sock->sk->sk_type != SOCK_RAW) {
1468		r = -ESOCKTNOSUPPORT;
1469		goto err;
1470	}
1471
1472	if (sock->sk->sk_family != AF_PACKET) {
1473		r = -EPFNOSUPPORT;
1474		goto err;
1475	}
1476	return sock;
1477err:
1478	sockfd_put(sock);
1479	return ERR_PTR(r);
1480}
1481
1482static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1483{
1484	struct ptr_ring *ring;
1485	ring = tun_get_tx_ring(file);
1486	if (!IS_ERR(ring))
1487		goto out;
1488	ring = tap_get_ptr_ring(file);
1489	if (!IS_ERR(ring))
1490		goto out;
1491	ring = NULL;
1492out:
1493	return ring;
1494}
1495
1496static struct socket *get_tap_socket(int fd)
1497{
1498	struct file *file = fget(fd);
1499	struct socket *sock;
1500
1501	if (!file)
1502		return ERR_PTR(-EBADF);
1503	sock = tun_get_socket(file);
1504	if (!IS_ERR(sock))
1505		return sock;
1506	sock = tap_get_socket(file);
1507	if (IS_ERR(sock))
1508		fput(file);
1509	return sock;
1510}
1511
1512static struct socket *get_socket(int fd)
1513{
1514	struct socket *sock;
1515
1516	/* special case to disable backend */
1517	if (fd == -1)
1518		return NULL;
1519	sock = get_raw_socket(fd);
1520	if (!IS_ERR(sock))
1521		return sock;
1522	sock = get_tap_socket(fd);
1523	if (!IS_ERR(sock))
1524		return sock;
1525	return ERR_PTR(-ENOTSOCK);
1526}
1527
1528static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1529{
1530	struct socket *sock, *oldsock;
1531	struct vhost_virtqueue *vq;
1532	struct vhost_net_virtqueue *nvq;
1533	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1534	int r;
1535
1536	mutex_lock(&n->dev.mutex);
1537	r = vhost_dev_check_owner(&n->dev);
1538	if (r)
1539		goto err;
1540
1541	if (index >= VHOST_NET_VQ_MAX) {
1542		r = -ENOBUFS;
1543		goto err;
1544	}
1545	vq = &n->vqs[index].vq;
1546	nvq = &n->vqs[index];
1547	mutex_lock(&vq->mutex);
1548
1549	if (fd == -1)
1550		vhost_clear_msg(&n->dev);
1551
1552	/* Verify that ring has been setup correctly. */
1553	if (!vhost_vq_access_ok(vq)) {
1554		r = -EFAULT;
1555		goto err_vq;
1556	}
1557	sock = get_socket(fd);
1558	if (IS_ERR(sock)) {
1559		r = PTR_ERR(sock);
1560		goto err_vq;
1561	}
1562
1563	/* start polling new socket */
1564	oldsock = vhost_vq_get_backend(vq);
1565	if (sock != oldsock) {
1566		ubufs = vhost_net_ubuf_alloc(vq,
1567					     sock && vhost_sock_zcopy(sock));
1568		if (IS_ERR(ubufs)) {
1569			r = PTR_ERR(ubufs);
1570			goto err_ubufs;
1571		}
1572
1573		vhost_net_disable_vq(n, vq);
1574		vhost_vq_set_backend(vq, sock);
1575		vhost_net_buf_unproduce(nvq);
1576		r = vhost_vq_init_access(vq);
1577		if (r)
1578			goto err_used;
1579		r = vhost_net_enable_vq(n, vq);
1580		if (r)
1581			goto err_used;
1582		if (index == VHOST_NET_VQ_RX) {
1583			if (sock)
1584				nvq->rx_ring = get_tap_ptr_ring(sock->file);
1585			else
1586				nvq->rx_ring = NULL;
1587		}
1588
1589		oldubufs = nvq->ubufs;
1590		nvq->ubufs = ubufs;
1591
1592		n->tx_packets = 0;
1593		n->tx_zcopy_err = 0;
1594		n->tx_flush = false;
1595	}
1596
1597	mutex_unlock(&vq->mutex);
1598
1599	if (oldubufs) {
1600		vhost_net_ubuf_put_wait_and_free(oldubufs);
1601		mutex_lock(&vq->mutex);
1602		vhost_zerocopy_signal_used(n, vq);
1603		mutex_unlock(&vq->mutex);
1604	}
1605
1606	if (oldsock) {
1607		vhost_dev_flush(&n->dev);
1608		sockfd_put(oldsock);
1609	}
1610
1611	mutex_unlock(&n->dev.mutex);
1612	return 0;
1613
1614err_used:
1615	vhost_vq_set_backend(vq, oldsock);
1616	vhost_net_enable_vq(n, vq);
1617	if (ubufs)
1618		vhost_net_ubuf_put_wait_and_free(ubufs);
1619err_ubufs:
1620	if (sock)
1621		sockfd_put(sock);
1622err_vq:
1623	mutex_unlock(&vq->mutex);
1624err:
1625	mutex_unlock(&n->dev.mutex);
1626	return r;
1627}
1628
1629static long vhost_net_reset_owner(struct vhost_net *n)
1630{
1631	struct socket *tx_sock = NULL;
1632	struct socket *rx_sock = NULL;
1633	long err;
1634	struct vhost_iotlb *umem;
1635
1636	mutex_lock(&n->dev.mutex);
1637	err = vhost_dev_check_owner(&n->dev);
1638	if (err)
1639		goto done;
1640	umem = vhost_dev_reset_owner_prepare();
1641	if (!umem) {
1642		err = -ENOMEM;
1643		goto done;
1644	}
1645	vhost_net_stop(n, &tx_sock, &rx_sock);
1646	vhost_net_flush(n);
1647	vhost_dev_stop(&n->dev);
1648	vhost_dev_reset_owner(&n->dev, umem);
1649	vhost_net_vq_reset(n);
1650done:
1651	mutex_unlock(&n->dev.mutex);
1652	if (tx_sock)
1653		sockfd_put(tx_sock);
1654	if (rx_sock)
1655		sockfd_put(rx_sock);
1656	return err;
1657}
1658
1659static int vhost_net_set_features(struct vhost_net *n, const u64 *features)
1660{
1661	size_t vhost_hlen, sock_hlen, hdr_len;
1662	int i;
1663
1664	hdr_len = virtio_features_test_bit(features, VIRTIO_NET_F_MRG_RXBUF) ||
1665		  virtio_features_test_bit(features, VIRTIO_F_VERSION_1) ?
1666		  sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1667		  sizeof(struct virtio_net_hdr);
1668
1669	if (virtio_features_test_bit(features,
1670				     VIRTIO_NET_F_HOST_UDP_TUNNEL_GSO) ||
1671	    virtio_features_test_bit(features,
1672				     VIRTIO_NET_F_GUEST_UDP_TUNNEL_GSO))
1673		hdr_len = sizeof(struct virtio_net_hdr_v1_hash_tunnel);
1674
1675	if (virtio_features_test_bit(features, VHOST_NET_F_VIRTIO_NET_HDR)) {
1676		/* vhost provides vnet_hdr */
1677		vhost_hlen = hdr_len;
1678		sock_hlen = 0;
1679	} else {
1680		/* socket provides vnet_hdr */
1681		vhost_hlen = 0;
1682		sock_hlen = hdr_len;
1683	}
1684	mutex_lock(&n->dev.mutex);
1685	if (virtio_features_test_bit(features, VHOST_F_LOG_ALL) &&
1686	    !vhost_log_access_ok(&n->dev))
1687		goto out_unlock;
1688
1689	if (virtio_features_test_bit(features, VIRTIO_F_ACCESS_PLATFORM)) {
1690		if (vhost_init_device_iotlb(&n->dev))
1691			goto out_unlock;
1692	}
1693
1694	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1695		mutex_lock(&n->vqs[i].vq.mutex);
1696		virtio_features_copy(n->vqs[i].vq.acked_features_array,
1697				     features);
1698		n->vqs[i].vhost_hlen = vhost_hlen;
1699		n->vqs[i].sock_hlen = sock_hlen;
1700		mutex_unlock(&n->vqs[i].vq.mutex);
1701	}
1702	mutex_unlock(&n->dev.mutex);
1703	return 0;
1704
1705out_unlock:
1706	mutex_unlock(&n->dev.mutex);
1707	return -EFAULT;
1708}
1709
1710static long vhost_net_set_owner(struct vhost_net *n)
1711{
1712	int r;
1713
1714	mutex_lock(&n->dev.mutex);
1715	if (vhost_dev_has_owner(&n->dev)) {
1716		r = -EBUSY;
1717		goto out;
1718	}
1719	r = vhost_net_set_ubuf_info(n);
1720	if (r)
1721		goto out;
1722	r = vhost_dev_set_owner(&n->dev);
1723	if (r)
1724		vhost_net_clear_ubuf_info(n);
1725	vhost_net_flush(n);
1726out:
1727	mutex_unlock(&n->dev.mutex);
1728	return r;
1729}
1730
1731static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1732			    unsigned long arg)
1733{
1734	const DEFINE_VHOST_FEATURES_ARRAY(vhost_net_features, vhost_net_bits);
1735	u64 all_features[VIRTIO_FEATURES_U64S];
1736	struct vhost_net *n = f->private_data;
1737	void __user *argp = (void __user *)arg;
1738	u64 __user *featurep = argp;
1739	struct vhost_vring_file backend;
1740	u64 features, count, copied;
1741	int r, i;
1742
1743	switch (ioctl) {
1744	case VHOST_NET_SET_BACKEND:
1745		if (copy_from_user(&backend, argp, sizeof backend))
1746			return -EFAULT;
1747		return vhost_net_set_backend(n, backend.index, backend.fd);
1748	case VHOST_GET_FEATURES:
1749		features = vhost_net_features[0];
1750		if (copy_to_user(featurep, &features, sizeof features))
1751			return -EFAULT;
1752		return 0;
1753	case VHOST_SET_FEATURES:
1754		if (copy_from_user(&features, featurep, sizeof features))
1755			return -EFAULT;
1756		if (features & ~vhost_net_features[0])
1757			return -EOPNOTSUPP;
1758
1759		virtio_features_from_u64(all_features, features);
1760		return vhost_net_set_features(n, all_features);
1761	case VHOST_GET_FEATURES_ARRAY:
1762		if (copy_from_user(&count, featurep, sizeof(count)))
1763			return -EFAULT;
1764
1765		/* Copy the net features, up to the user-provided buffer size */
1766		argp += sizeof(u64);
1767		copied = min(count, (u64)VIRTIO_FEATURES_U64S);
1768		if (copy_to_user(argp, vhost_net_features,
1769				 copied * sizeof(u64)))
1770			return -EFAULT;
1771
1772		/* Zero the trailing space provided by user-space, if any */
1773		if (clear_user(argp, size_mul(count - copied, sizeof(u64))))
1774			return -EFAULT;
1775		return 0;
1776	case VHOST_SET_FEATURES_ARRAY:
1777		if (copy_from_user(&count, featurep, sizeof(count)))
1778			return -EFAULT;
1779
1780		virtio_features_zero(all_features);
1781		argp += sizeof(u64);
1782		copied = min(count, (u64)VIRTIO_FEATURES_U64S);
1783		if (copy_from_user(all_features, argp, copied * sizeof(u64)))
1784			return -EFAULT;
1785
1786		/*
1787		 * Any feature specified by user-space above
1788		 * VIRTIO_FEATURES_BITS is not supported by definition.
1789		 */
1790		for (i = copied; i < count; ++i) {
1791			if (copy_from_user(&features, featurep + 1 + i,
1792					   sizeof(features)))
1793				return -EFAULT;
1794			if (features)
1795				return -EOPNOTSUPP;
1796		}
1797
1798		for (i = 0; i < VIRTIO_FEATURES_U64S; i++)
1799			if (all_features[i] & ~vhost_net_features[i])
1800				return -EOPNOTSUPP;
1801
1802		return vhost_net_set_features(n, all_features);
1803	case VHOST_GET_BACKEND_FEATURES:
1804		features = VHOST_NET_BACKEND_FEATURES;
1805		if (copy_to_user(featurep, &features, sizeof(features)))
1806			return -EFAULT;
1807		return 0;
1808	case VHOST_SET_BACKEND_FEATURES:
1809		if (copy_from_user(&features, featurep, sizeof(features)))
1810			return -EFAULT;
1811		if (features & ~VHOST_NET_BACKEND_FEATURES)
1812			return -EOPNOTSUPP;
1813		vhost_set_backend_features(&n->dev, features);
1814		return 0;
1815	case VHOST_RESET_OWNER:
1816		return vhost_net_reset_owner(n);
1817	case VHOST_SET_OWNER:
1818		return vhost_net_set_owner(n);
1819	default:
1820		mutex_lock(&n->dev.mutex);
1821		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1822		if (r == -ENOIOCTLCMD)
1823			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1824		else
1825			vhost_net_flush(n);
1826		mutex_unlock(&n->dev.mutex);
1827		return r;
1828	}
1829}
1830
1831static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1832{
1833	struct file *file = iocb->ki_filp;
1834	struct vhost_net *n = file->private_data;
1835	struct vhost_dev *dev = &n->dev;
1836	int noblock = file->f_flags & O_NONBLOCK;
1837
1838	return vhost_chr_read_iter(dev, to, noblock);
1839}
1840
1841static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1842					struct iov_iter *from)
1843{
1844	struct file *file = iocb->ki_filp;
1845	struct vhost_net *n = file->private_data;
1846	struct vhost_dev *dev = &n->dev;
1847
1848	return vhost_chr_write_iter(dev, from);
1849}
1850
1851static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1852{
1853	struct vhost_net *n = file->private_data;
1854	struct vhost_dev *dev = &n->dev;
1855
1856	return vhost_chr_poll(file, dev, wait);
1857}
1858
1859static const struct file_operations vhost_net_fops = {
1860	.owner          = THIS_MODULE,
1861	.release        = vhost_net_release,
1862	.read_iter      = vhost_net_chr_read_iter,
1863	.write_iter     = vhost_net_chr_write_iter,
1864	.poll           = vhost_net_chr_poll,
1865	.unlocked_ioctl = vhost_net_ioctl,
1866	.compat_ioctl   = compat_ptr_ioctl,
1867	.open           = vhost_net_open,
1868	.llseek		= noop_llseek,
1869};
1870
1871static struct miscdevice vhost_net_misc = {
1872	.minor = VHOST_NET_MINOR,
1873	.name = "vhost-net",
1874	.fops = &vhost_net_fops,
1875};
1876
1877static int __init vhost_net_init(void)
1878{
1879	if (experimental_zcopytx)
1880		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1881	return misc_register(&vhost_net_misc);
1882}
1883module_init(vhost_net_init);
1884
1885static void __exit vhost_net_exit(void)
1886{
1887	misc_deregister(&vhost_net_misc);
1888}
1889module_exit(vhost_net_exit);
1890
1891MODULE_VERSION("0.0.1");
1892MODULE_LICENSE("GPL v2");
1893MODULE_AUTHOR("Michael S. Tsirkin");
1894MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1895MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1896MODULE_ALIAS("devname:vhost-net");