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
  72enum {
  73	VHOST_NET_FEATURES = VHOST_FEATURES |
  74			 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  75			 (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  76			 (1ULL << VIRTIO_F_ACCESS_PLATFORM)
  77};
  78
  79enum {
  80	VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  81};
  82
  83enum {
  84	VHOST_NET_VQ_RX = 0,
  85	VHOST_NET_VQ_TX = 1,
  86	VHOST_NET_VQ_MAX = 2,
  87};
  88
  89struct vhost_net_ubuf_ref {
  90	/* refcount follows semantics similar to kref:
  91	 *  0: object is released
  92	 *  1: no outstanding ubufs
  93	 * >1: outstanding ubufs
  94	 */
  95	atomic_t refcount;
  96	wait_queue_head_t wait;
  97	struct vhost_virtqueue *vq;
  98};
  99
 100#define VHOST_NET_BATCH 64
 101struct vhost_net_buf {
 102	void **queue;
 103	int tail;
 104	int head;
 105};
 106
 107struct vhost_net_virtqueue {
 108	struct vhost_virtqueue vq;
 109	size_t vhost_hlen;
 110	size_t sock_hlen;
 111	/* vhost zerocopy support fields below: */
 112	/* last used idx for outstanding DMA zerocopy buffers */
 113	int upend_idx;
 114	/* For TX, first used idx for DMA done zerocopy buffers
 115	 * For RX, number of batched heads
 116	 */
 117	int done_idx;
 118	/* Number of XDP frames batched */
 119	int batched_xdp;
 120	/* an array of userspace buffers info */
 121	struct ubuf_info *ubuf_info;
 122	/* Reference counting for outstanding ubufs.
 123	 * Protected by vq mutex. Writers must also take device mutex. */
 124	struct vhost_net_ubuf_ref *ubufs;
 125	struct ptr_ring *rx_ring;
 126	struct vhost_net_buf rxq;
 127	/* Batched XDP buffs */
 128	struct xdp_buff *xdp;
 129};
 130
 131struct vhost_net {
 132	struct vhost_dev dev;
 133	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 134	struct vhost_poll poll[VHOST_NET_VQ_MAX];
 135	/* Number of TX recently submitted.
 136	 * Protected by tx vq lock. */
 137	unsigned tx_packets;
 138	/* Number of times zerocopy TX recently failed.
 139	 * Protected by tx vq lock. */
 140	unsigned tx_zcopy_err;
 141	/* Flush in progress. Protected by tx vq lock. */
 142	bool tx_flush;
 143	/* Private page frag */
 144	struct page_frag page_frag;
 145	/* Refcount bias of page frag */
 146	int refcnt_bias;
 147};
 148
 149static unsigned vhost_net_zcopy_mask __read_mostly;
 150
 151static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 152{
 153	if (rxq->tail != rxq->head)
 154		return rxq->queue[rxq->head];
 155	else
 156		return NULL;
 157}
 158
 159static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 160{
 161	return rxq->tail - rxq->head;
 162}
 163
 164static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 165{
 166	return rxq->tail == rxq->head;
 167}
 168
 169static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 170{
 171	void *ret = vhost_net_buf_get_ptr(rxq);
 172	++rxq->head;
 173	return ret;
 174}
 175
 176static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 177{
 178	struct vhost_net_buf *rxq = &nvq->rxq;
 179
 180	rxq->head = 0;
 181	rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 182					      VHOST_NET_BATCH);
 183	return rxq->tail;
 184}
 185
 186static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 187{
 188	struct vhost_net_buf *rxq = &nvq->rxq;
 189
 190	if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 191		ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 192				   vhost_net_buf_get_size(rxq),
 193				   tun_ptr_free);
 194		rxq->head = rxq->tail = 0;
 195	}
 196}
 197
 198static int vhost_net_buf_peek_len(void *ptr)
 199{
 200	if (tun_is_xdp_frame(ptr)) {
 201		struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 202
 203		return xdpf->len;
 204	}
 205
 206	return __skb_array_len_with_tag(ptr);
 207}
 208
 209static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 210{
 211	struct vhost_net_buf *rxq = &nvq->rxq;
 212
 213	if (!vhost_net_buf_is_empty(rxq))
 214		goto out;
 215
 216	if (!vhost_net_buf_produce(nvq))
 217		return 0;
 218
 219out:
 220	return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 221}
 222
 223static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 224{
 225	rxq->head = rxq->tail = 0;
 226}
 227
 228static void vhost_net_enable_zcopy(int vq)
 229{
 230	vhost_net_zcopy_mask |= 0x1 << vq;
 231}
 232
 233static struct vhost_net_ubuf_ref *
 234vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 235{
 236	struct vhost_net_ubuf_ref *ubufs;
 237	/* No zero copy backend? Nothing to count. */
 238	if (!zcopy)
 239		return NULL;
 240	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 241	if (!ubufs)
 242		return ERR_PTR(-ENOMEM);
 243	atomic_set(&ubufs->refcount, 1);
 244	init_waitqueue_head(&ubufs->wait);
 245	ubufs->vq = vq;
 246	return ubufs;
 247}
 248
 249static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 250{
 251	int r = atomic_sub_return(1, &ubufs->refcount);
 252	if (unlikely(!r))
 253		wake_up(&ubufs->wait);
 254	return r;
 255}
 256
 257static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 258{
 259	vhost_net_ubuf_put(ubufs);
 260	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 261}
 262
 263static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 264{
 265	vhost_net_ubuf_put_and_wait(ubufs);
 266	kfree(ubufs);
 267}
 268
 269static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 270{
 271	int i;
 272
 273	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 274		kfree(n->vqs[i].ubuf_info);
 275		n->vqs[i].ubuf_info = NULL;
 276	}
 277}
 278
 279static int vhost_net_set_ubuf_info(struct vhost_net *n)
 280{
 281	bool zcopy;
 282	int i;
 283
 284	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 285		zcopy = vhost_net_zcopy_mask & (0x1 << i);
 286		if (!zcopy)
 287			continue;
 288		n->vqs[i].ubuf_info =
 289			kmalloc_array(UIO_MAXIOV,
 290				      sizeof(*n->vqs[i].ubuf_info),
 291				      GFP_KERNEL);
 292		if  (!n->vqs[i].ubuf_info)
 293			goto err;
 294	}
 295	return 0;
 296
 297err:
 298	vhost_net_clear_ubuf_info(n);
 299	return -ENOMEM;
 300}
 301
 302static void vhost_net_vq_reset(struct vhost_net *n)
 303{
 304	int i;
 305
 306	vhost_net_clear_ubuf_info(n);
 307
 308	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 309		n->vqs[i].done_idx = 0;
 310		n->vqs[i].upend_idx = 0;
 311		n->vqs[i].ubufs = NULL;
 312		n->vqs[i].vhost_hlen = 0;
 313		n->vqs[i].sock_hlen = 0;
 314		vhost_net_buf_init(&n->vqs[i].rxq);
 315	}
 316
 317}
 318
 319static void vhost_net_tx_packet(struct vhost_net *net)
 320{
 321	++net->tx_packets;
 322	if (net->tx_packets < 1024)
 323		return;
 324	net->tx_packets = 0;
 325	net->tx_zcopy_err = 0;
 326}
 327
 328static void vhost_net_tx_err(struct vhost_net *net)
 329{
 330	++net->tx_zcopy_err;
 331}
 332
 333static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 334{
 335	/* TX flush waits for outstanding DMAs to be done.
 336	 * Don't start new DMAs.
 337	 */
 338	return !net->tx_flush &&
 339		net->tx_packets / 64 >= net->tx_zcopy_err;
 340}
 341
 342static bool vhost_sock_zcopy(struct socket *sock)
 343{
 344	return unlikely(experimental_zcopytx) &&
 345		sock_flag(sock->sk, SOCK_ZEROCOPY);
 346}
 347
 348static bool vhost_sock_xdp(struct socket *sock)
 349{
 350	return sock_flag(sock->sk, SOCK_XDP);
 351}
 352
 353/* In case of DMA done not in order in lower device driver for some reason.
 354 * upend_idx is used to track end of used idx, done_idx is used to track head
 355 * of used idx. Once lower device DMA done contiguously, we will signal KVM
 356 * guest used idx.
 357 */
 358static void vhost_zerocopy_signal_used(struct vhost_net *net,
 359				       struct vhost_virtqueue *vq)
 360{
 361	struct vhost_net_virtqueue *nvq =
 362		container_of(vq, struct vhost_net_virtqueue, vq);
 363	int i, add;
 364	int j = 0;
 365
 366	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 367		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 368			vhost_net_tx_err(net);
 369		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 370			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 371			++j;
 372		} else
 373			break;
 374	}
 375	while (j) {
 376		add = min(UIO_MAXIOV - nvq->done_idx, j);
 377		vhost_add_used_and_signal_n(vq->dev, vq,
 378					    &vq->heads[nvq->done_idx], add);
 379		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 380		j -= add;
 381	}
 382}
 383
 384static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 385{
 386	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 387	struct vhost_virtqueue *vq = ubufs->vq;
 388	int cnt;
 389
 390	rcu_read_lock_bh();
 391
 392	/* set len to mark this desc buffers done DMA */
 393	vq->heads[ubuf->desc].len = success ?
 394		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 395	cnt = vhost_net_ubuf_put(ubufs);
 396
 397	/*
 398	 * Trigger polling thread if guest stopped submitting new buffers:
 399	 * in this case, the refcount after decrement will eventually reach 1.
 400	 * We also trigger polling periodically after each 16 packets
 401	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
 402	 * less than 10% of times).
 403	 */
 404	if (cnt <= 1 || !(cnt % 16))
 405		vhost_poll_queue(&vq->poll);
 406
 407	rcu_read_unlock_bh();
 408}
 409
 410static inline unsigned long busy_clock(void)
 411{
 412	return local_clock() >> 10;
 413}
 414
 415static bool vhost_can_busy_poll(unsigned long endtime)
 416{
 417	return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 418		      !signal_pending(current));
 419}
 420
 421static void vhost_net_disable_vq(struct vhost_net *n,
 422				 struct vhost_virtqueue *vq)
 423{
 424	struct vhost_net_virtqueue *nvq =
 425		container_of(vq, struct vhost_net_virtqueue, vq);
 426	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 427	if (!vhost_vq_get_backend(vq))
 428		return;
 429	vhost_poll_stop(poll);
 430}
 431
 432static int vhost_net_enable_vq(struct vhost_net *n,
 433				struct vhost_virtqueue *vq)
 434{
 435	struct vhost_net_virtqueue *nvq =
 436		container_of(vq, struct vhost_net_virtqueue, vq);
 437	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 438	struct socket *sock;
 439
 440	sock = vhost_vq_get_backend(vq);
 441	if (!sock)
 442		return 0;
 443
 444	return vhost_poll_start(poll, sock->file);
 445}
 446
 447static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 448{
 449	struct vhost_virtqueue *vq = &nvq->vq;
 450	struct vhost_dev *dev = vq->dev;
 451
 452	if (!nvq->done_idx)
 453		return;
 454
 455	vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 456	nvq->done_idx = 0;
 457}
 458
 459static void vhost_tx_batch(struct vhost_net *net,
 460			   struct vhost_net_virtqueue *nvq,
 461			   struct socket *sock,
 462			   struct msghdr *msghdr)
 463{
 464	struct tun_msg_ctl ctl = {
 465		.type = TUN_MSG_PTR,
 466		.num = nvq->batched_xdp,
 467		.ptr = nvq->xdp,
 468	};
 469	int err;
 470
 471	if (nvq->batched_xdp == 0)
 472		goto signal_used;
 473
 474	msghdr->msg_control = &ctl;
 475	err = sock->ops->sendmsg(sock, msghdr, 0);
 476	if (unlikely(err < 0)) {
 477		vq_err(&nvq->vq, "Fail to batch sending packets\n");
 478		return;
 479	}
 480
 481signal_used:
 482	vhost_net_signal_used(nvq);
 483	nvq->batched_xdp = 0;
 484}
 485
 486static int sock_has_rx_data(struct socket *sock)
 487{
 488	if (unlikely(!sock))
 489		return 0;
 490
 491	if (sock->ops->peek_len)
 492		return sock->ops->peek_len(sock);
 493
 494	return skb_queue_empty(&sock->sk->sk_receive_queue);
 495}
 496
 497static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
 498					  struct vhost_virtqueue *vq)
 499{
 500	if (!vhost_vq_avail_empty(&net->dev, vq)) {
 501		vhost_poll_queue(&vq->poll);
 502	} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 503		vhost_disable_notify(&net->dev, vq);
 504		vhost_poll_queue(&vq->poll);
 505	}
 506}
 507
 508static void vhost_net_busy_poll(struct vhost_net *net,
 509				struct vhost_virtqueue *rvq,
 510				struct vhost_virtqueue *tvq,
 511				bool *busyloop_intr,
 512				bool poll_rx)
 513{
 514	unsigned long busyloop_timeout;
 515	unsigned long endtime;
 516	struct socket *sock;
 517	struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
 518
 519	/* Try to hold the vq mutex of the paired virtqueue. We can't
 520	 * use mutex_lock() here since we could not guarantee a
 521	 * consistenet lock ordering.
 522	 */
 523	if (!mutex_trylock(&vq->mutex))
 524		return;
 525
 526	vhost_disable_notify(&net->dev, vq);
 527	sock = vhost_vq_get_backend(rvq);
 528
 529	busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
 530				     tvq->busyloop_timeout;
 531
 532	preempt_disable();
 533	endtime = busy_clock() + busyloop_timeout;
 534
 535	while (vhost_can_busy_poll(endtime)) {
 536		if (vhost_has_work(&net->dev)) {
 537			*busyloop_intr = true;
 538			break;
 539		}
 540
 541		if ((sock_has_rx_data(sock) &&
 542		     !vhost_vq_avail_empty(&net->dev, rvq)) ||
 543		    !vhost_vq_avail_empty(&net->dev, tvq))
 544			break;
 545
 546		cpu_relax();
 547	}
 548
 549	preempt_enable();
 550
 551	if (poll_rx || sock_has_rx_data(sock))
 552		vhost_net_busy_poll_try_queue(net, vq);
 553	else if (!poll_rx) /* On tx here, sock has no rx data. */
 554		vhost_enable_notify(&net->dev, rvq);
 555
 556	mutex_unlock(&vq->mutex);
 557}
 558
 559static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 560				    struct vhost_net_virtqueue *tnvq,
 561				    unsigned int *out_num, unsigned int *in_num,
 562				    struct msghdr *msghdr, bool *busyloop_intr)
 563{
 564	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 565	struct vhost_virtqueue *rvq = &rnvq->vq;
 566	struct vhost_virtqueue *tvq = &tnvq->vq;
 567
 568	int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 569				  out_num, in_num, NULL, NULL);
 570
 571	if (r == tvq->num && tvq->busyloop_timeout) {
 572		/* Flush batched packets first */
 573		if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
 574			vhost_tx_batch(net, tnvq,
 575				       vhost_vq_get_backend(tvq),
 576				       msghdr);
 577
 578		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
 579
 580		r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
 581				      out_num, in_num, NULL, NULL);
 582	}
 583
 584	return r;
 585}
 586
 587static bool vhost_exceeds_maxpend(struct vhost_net *net)
 588{
 589	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 590	struct vhost_virtqueue *vq = &nvq->vq;
 591
 592	return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 593	       min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 594}
 595
 596static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 597			    size_t hdr_size, int out)
 598{
 599	/* Skip header. TODO: support TSO. */
 600	size_t len = iov_length(vq->iov, out);
 601
 602	iov_iter_init(iter, WRITE, vq->iov, out, len);
 603	iov_iter_advance(iter, hdr_size);
 604
 605	return iov_iter_count(iter);
 606}
 607
 608static int get_tx_bufs(struct vhost_net *net,
 609		       struct vhost_net_virtqueue *nvq,
 610		       struct msghdr *msg,
 611		       unsigned int *out, unsigned int *in,
 612		       size_t *len, bool *busyloop_intr)
 613{
 614	struct vhost_virtqueue *vq = &nvq->vq;
 615	int ret;
 616
 617	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
 618
 619	if (ret < 0 || ret == vq->num)
 620		return ret;
 621
 622	if (*in) {
 623		vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 624			*out, *in);
 625		return -EFAULT;
 626	}
 627
 628	/* Sanity check */
 629	*len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 630	if (*len == 0) {
 631		vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 632			*len, nvq->vhost_hlen);
 633		return -EFAULT;
 634	}
 635
 636	return ret;
 637}
 638
 639static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 640{
 641	return total_len < VHOST_NET_WEIGHT &&
 642	       !vhost_vq_avail_empty(vq->dev, vq);
 643}
 644
 645#define SKB_FRAG_PAGE_ORDER     get_order(32768)
 646
 647static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
 648				       struct page_frag *pfrag, gfp_t gfp)
 649{
 650	if (pfrag->page) {
 651		if (pfrag->offset + sz <= pfrag->size)
 652			return true;
 653		__page_frag_cache_drain(pfrag->page, net->refcnt_bias);
 654	}
 655
 656	pfrag->offset = 0;
 657	net->refcnt_bias = 0;
 658	if (SKB_FRAG_PAGE_ORDER) {
 659		/* Avoid direct reclaim but allow kswapd to wake */
 660		pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
 661					  __GFP_COMP | __GFP_NOWARN |
 662					  __GFP_NORETRY,
 663					  SKB_FRAG_PAGE_ORDER);
 664		if (likely(pfrag->page)) {
 665			pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
 666			goto done;
 667		}
 668	}
 669	pfrag->page = alloc_page(gfp);
 670	if (likely(pfrag->page)) {
 671		pfrag->size = PAGE_SIZE;
 672		goto done;
 673	}
 674	return false;
 675
 676done:
 677	net->refcnt_bias = USHRT_MAX;
 678	page_ref_add(pfrag->page, USHRT_MAX - 1);
 679	return true;
 680}
 681
 682#define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 683
 684static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 685			       struct iov_iter *from)
 686{
 687	struct vhost_virtqueue *vq = &nvq->vq;
 688	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
 689					     dev);
 690	struct socket *sock = vhost_vq_get_backend(vq);
 691	struct page_frag *alloc_frag = &net->page_frag;
 692	struct virtio_net_hdr *gso;
 693	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 694	struct tun_xdp_hdr *hdr;
 695	size_t len = iov_iter_count(from);
 696	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
 697	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 698	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
 699	int sock_hlen = nvq->sock_hlen;
 700	void *buf;
 701	int copied;
 702
 703	if (unlikely(len < nvq->sock_hlen))
 704		return -EFAULT;
 705
 706	if (SKB_DATA_ALIGN(len + pad) +
 707	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
 708		return -ENOSPC;
 709
 710	buflen += SKB_DATA_ALIGN(len + pad);
 711	alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
 712	if (unlikely(!vhost_net_page_frag_refill(net, buflen,
 713						 alloc_frag, GFP_KERNEL)))
 714		return -ENOMEM;
 715
 716	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 717	copied = copy_page_from_iter(alloc_frag->page,
 718				     alloc_frag->offset +
 719				     offsetof(struct tun_xdp_hdr, gso),
 720				     sock_hlen, from);
 721	if (copied != sock_hlen)
 722		return -EFAULT;
 723
 724	hdr = buf;
 725	gso = &hdr->gso;
 726
 727	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 728	    vhost16_to_cpu(vq, gso->csum_start) +
 729	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
 730	    vhost16_to_cpu(vq, gso->hdr_len)) {
 731		gso->hdr_len = cpu_to_vhost16(vq,
 732			       vhost16_to_cpu(vq, gso->csum_start) +
 733			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
 734
 735		if (vhost16_to_cpu(vq, gso->hdr_len) > len)
 736			return -EINVAL;
 737	}
 738
 739	len -= sock_hlen;
 740	copied = copy_page_from_iter(alloc_frag->page,
 741				     alloc_frag->offset + pad,
 742				     len, from);
 743	if (copied != len)
 744		return -EFAULT;
 745
 746	xdp->data_hard_start = buf;
 747	xdp->data = buf + pad;
 748	xdp->data_end = xdp->data + len;
 749	hdr->buflen = buflen;
 750	xdp->frame_sz = buflen;
 751
 752	--net->refcnt_bias;
 753	alloc_frag->offset += buflen;
 754
 755	++nvq->batched_xdp;
 756
 757	return 0;
 758}
 759
 760static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 761{
 762	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 763	struct vhost_virtqueue *vq = &nvq->vq;
 764	unsigned out, in;
 765	int head;
 766	struct msghdr msg = {
 767		.msg_name = NULL,
 768		.msg_namelen = 0,
 769		.msg_control = NULL,
 770		.msg_controllen = 0,
 771		.msg_flags = MSG_DONTWAIT,
 772	};
 773	size_t len, total_len = 0;
 774	int err;
 775	int sent_pkts = 0;
 776	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
 777
 778	do {
 779		bool busyloop_intr = false;
 780
 781		if (nvq->done_idx == VHOST_NET_BATCH)
 782			vhost_tx_batch(net, nvq, sock, &msg);
 783
 784		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 785				   &busyloop_intr);
 786		/* On error, stop handling until the next kick. */
 787		if (unlikely(head < 0))
 788			break;
 789		/* Nothing new?  Wait for eventfd to tell us they refilled. */
 790		if (head == vq->num) {
 791			if (unlikely(busyloop_intr)) {
 792				vhost_poll_queue(&vq->poll);
 793			} else if (unlikely(vhost_enable_notify(&net->dev,
 794								vq))) {
 795				vhost_disable_notify(&net->dev, vq);
 796				continue;
 797			}
 798			break;
 799		}
 800
 801		total_len += len;
 802
 803		/* For simplicity, TX batching is only enabled if
 804		 * sndbuf is unlimited.
 805		 */
 806		if (sock_can_batch) {
 807			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
 808			if (!err) {
 809				goto done;
 810			} else if (unlikely(err != -ENOSPC)) {
 811				vhost_tx_batch(net, nvq, sock, &msg);
 812				vhost_discard_vq_desc(vq, 1);
 813				vhost_net_enable_vq(net, vq);
 814				break;
 815			}
 816
 817			/* We can't build XDP buff, go for single
 818			 * packet path but let's flush batched
 819			 * packets.
 820			 */
 821			vhost_tx_batch(net, nvq, sock, &msg);
 822			msg.msg_control = NULL;
 823		} else {
 824			if (tx_can_batch(vq, total_len))
 825				msg.msg_flags |= MSG_MORE;
 826			else
 827				msg.msg_flags &= ~MSG_MORE;
 828		}
 829
 830		/* TODO: Check specific error and bomb out unless ENOBUFS? */
 831		err = sock->ops->sendmsg(sock, &msg, len);
 832		if (unlikely(err < 0)) {
 833			vhost_discard_vq_desc(vq, 1);
 834			vhost_net_enable_vq(net, vq);
 835			break;
 836		}
 837		if (err != len)
 838			pr_debug("Truncated TX packet: len %d != %zd\n",
 839				 err, len);
 840done:
 841		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 842		vq->heads[nvq->done_idx].len = 0;
 843		++nvq->done_idx;
 844	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 845
 846	vhost_tx_batch(net, nvq, sock, &msg);
 847}
 848
 849static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 850{
 851	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 852	struct vhost_virtqueue *vq = &nvq->vq;
 853	unsigned out, in;
 854	int head;
 855	struct msghdr msg = {
 856		.msg_name = NULL,
 857		.msg_namelen = 0,
 858		.msg_control = NULL,
 859		.msg_controllen = 0,
 860		.msg_flags = MSG_DONTWAIT,
 861	};
 862	struct tun_msg_ctl ctl;
 863	size_t len, total_len = 0;
 864	int err;
 865	struct vhost_net_ubuf_ref *ubufs;
 866	bool zcopy_used;
 867	int sent_pkts = 0;
 868
 869	do {
 870		bool busyloop_intr;
 871
 872		/* Release DMAs done buffers first */
 873		vhost_zerocopy_signal_used(net, vq);
 874
 875		busyloop_intr = false;
 876		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 877				   &busyloop_intr);
 878		/* On error, stop handling until the next kick. */
 879		if (unlikely(head < 0))
 880			break;
 881		/* Nothing new?  Wait for eventfd to tell us they refilled. */
 882		if (head == vq->num) {
 883			if (unlikely(busyloop_intr)) {
 884				vhost_poll_queue(&vq->poll);
 885			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 886				vhost_disable_notify(&net->dev, vq);
 887				continue;
 888			}
 889			break;
 890		}
 891
 892		zcopy_used = len >= VHOST_GOODCOPY_LEN
 893			     && !vhost_exceeds_maxpend(net)
 894			     && vhost_net_tx_select_zcopy(net);
 895
 896		/* use msg_control to pass vhost zerocopy ubuf info to skb */
 897		if (zcopy_used) {
 898			struct ubuf_info *ubuf;
 899			ubuf = nvq->ubuf_info + nvq->upend_idx;
 900
 901			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 902			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 903			ubuf->callback = vhost_zerocopy_callback;
 904			ubuf->ctx = nvq->ubufs;
 905			ubuf->desc = nvq->upend_idx;
 906			refcount_set(&ubuf->refcnt, 1);
 907			msg.msg_control = &ctl;
 908			ctl.type = TUN_MSG_UBUF;
 909			ctl.ptr = ubuf;
 910			msg.msg_controllen = sizeof(ctl);
 911			ubufs = nvq->ubufs;
 912			atomic_inc(&ubufs->refcount);
 913			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 914		} else {
 915			msg.msg_control = NULL;
 916			ubufs = NULL;
 917		}
 918		total_len += len;
 919		if (tx_can_batch(vq, total_len) &&
 920		    likely(!vhost_exceeds_maxpend(net))) {
 921			msg.msg_flags |= MSG_MORE;
 922		} else {
 923			msg.msg_flags &= ~MSG_MORE;
 924		}
 925
 926		/* TODO: Check specific error and bomb out unless ENOBUFS? */
 927		err = sock->ops->sendmsg(sock, &msg, len);
 928		if (unlikely(err < 0)) {
 929			if (zcopy_used) {
 930				vhost_net_ubuf_put(ubufs);
 931				nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 932					% UIO_MAXIOV;
 933			}
 934			vhost_discard_vq_desc(vq, 1);
 935			vhost_net_enable_vq(net, vq);
 936			break;
 937		}
 938		if (err != len)
 939			pr_debug("Truncated TX packet: "
 940				 " len %d != %zd\n", err, len);
 941		if (!zcopy_used)
 942			vhost_add_used_and_signal(&net->dev, vq, head, 0);
 943		else
 944			vhost_zerocopy_signal_used(net, vq);
 945		vhost_net_tx_packet(net);
 946	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 947}
 948
 949/* Expects to be always run from workqueue - which acts as
 950 * read-size critical section for our kind of RCU. */
 951static void handle_tx(struct vhost_net *net)
 952{
 953	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 954	struct vhost_virtqueue *vq = &nvq->vq;
 955	struct socket *sock;
 956
 957	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
 958	sock = vhost_vq_get_backend(vq);
 959	if (!sock)
 960		goto out;
 961
 962	if (!vq_meta_prefetch(vq))
 963		goto out;
 964
 965	vhost_disable_notify(&net->dev, vq);
 966	vhost_net_disable_vq(net, vq);
 967
 968	if (vhost_sock_zcopy(sock))
 969		handle_tx_zerocopy(net, sock);
 970	else
 971		handle_tx_copy(net, sock);
 972
 973out:
 974	mutex_unlock(&vq->mutex);
 975}
 976
 977static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 978{
 979	struct sk_buff *head;
 980	int len = 0;
 981	unsigned long flags;
 982
 983	if (rvq->rx_ring)
 984		return vhost_net_buf_peek(rvq);
 985
 986	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 987	head = skb_peek(&sk->sk_receive_queue);
 988	if (likely(head)) {
 989		len = head->len;
 990		if (skb_vlan_tag_present(head))
 991			len += VLAN_HLEN;
 992	}
 993
 994	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 995	return len;
 996}
 997
 998static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 999				      bool *busyloop_intr)
1000{
1001	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1002	struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1003	struct vhost_virtqueue *rvq = &rnvq->vq;
1004	struct vhost_virtqueue *tvq = &tnvq->vq;
1005	int len = peek_head_len(rnvq, sk);
1006
1007	if (!len && rvq->busyloop_timeout) {
1008		/* Flush batched heads first */
1009		vhost_net_signal_used(rnvq);
1010		/* Both tx vq and rx socket were polled here */
1011		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1012
1013		len = peek_head_len(rnvq, sk);
1014	}
1015
1016	return len;
1017}
1018
1019/* This is a multi-buffer version of vhost_get_desc, that works if
1020 *	vq has read descriptors only.
1021 * @vq		- the relevant virtqueue
1022 * @datalen	- data length we'll be reading
1023 * @iovcount	- returned count of io vectors we fill
1024 * @log		- vhost log
1025 * @log_num	- log offset
1026 * @quota       - headcount quota, 1 for big buffer
1027 *	returns number of buffer heads allocated, negative on error
1028 */
1029static int get_rx_bufs(struct vhost_virtqueue *vq,
1030		       struct vring_used_elem *heads,
1031		       int datalen,
1032		       unsigned *iovcount,
1033		       struct vhost_log *log,
1034		       unsigned *log_num,
1035		       unsigned int quota)
1036{
1037	unsigned int out, in;
1038	int seg = 0;
1039	int headcount = 0;
1040	unsigned d;
1041	int r, nlogs = 0;
1042	/* len is always initialized before use since we are always called with
1043	 * datalen > 0.
1044	 */
1045	u32 len;
1046
1047	while (datalen > 0 && headcount < quota) {
1048		if (unlikely(seg >= UIO_MAXIOV)) {
1049			r = -ENOBUFS;
1050			goto err;
1051		}
1052		r = vhost_get_vq_desc(vq, vq->iov + seg,
1053				      ARRAY_SIZE(vq->iov) - seg, &out,
1054				      &in, log, log_num);
1055		if (unlikely(r < 0))
1056			goto err;
1057
1058		d = r;
1059		if (d == vq->num) {
1060			r = 0;
1061			goto err;
1062		}
1063		if (unlikely(out || in <= 0)) {
1064			vq_err(vq, "unexpected descriptor format for RX: "
1065				"out %d, in %d\n", out, in);
1066			r = -EINVAL;
1067			goto err;
1068		}
1069		if (unlikely(log)) {
1070			nlogs += *log_num;
1071			log += *log_num;
1072		}
1073		heads[headcount].id = cpu_to_vhost32(vq, d);
1074		len = iov_length(vq->iov + seg, in);
1075		heads[headcount].len = cpu_to_vhost32(vq, len);
1076		datalen -= len;
1077		++headcount;
1078		seg += in;
1079	}
1080	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1081	*iovcount = seg;
1082	if (unlikely(log))
1083		*log_num = nlogs;
1084
1085	/* Detect overrun */
1086	if (unlikely(datalen > 0)) {
1087		r = UIO_MAXIOV + 1;
1088		goto err;
1089	}
1090	return headcount;
1091err:
1092	vhost_discard_vq_desc(vq, headcount);
1093	return r;
1094}
1095
1096/* Expects to be always run from workqueue - which acts as
1097 * read-size critical section for our kind of RCU. */
1098static void handle_rx(struct vhost_net *net)
1099{
1100	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1101	struct vhost_virtqueue *vq = &nvq->vq;
1102	unsigned in, log;
1103	struct vhost_log *vq_log;
1104	struct msghdr msg = {
1105		.msg_name = NULL,
1106		.msg_namelen = 0,
1107		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
1108		.msg_controllen = 0,
1109		.msg_flags = MSG_DONTWAIT,
1110	};
1111	struct virtio_net_hdr hdr = {
1112		.flags = 0,
1113		.gso_type = VIRTIO_NET_HDR_GSO_NONE
1114	};
1115	size_t total_len = 0;
1116	int err, mergeable;
1117	s16 headcount;
1118	size_t vhost_hlen, sock_hlen;
1119	size_t vhost_len, sock_len;
1120	bool busyloop_intr = false;
1121	struct socket *sock;
1122	struct iov_iter fixup;
1123	__virtio16 num_buffers;
1124	int recv_pkts = 0;
1125
1126	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1127	sock = vhost_vq_get_backend(vq);
1128	if (!sock)
1129		goto out;
1130
1131	if (!vq_meta_prefetch(vq))
1132		goto out;
1133
1134	vhost_disable_notify(&net->dev, vq);
1135	vhost_net_disable_vq(net, vq);
1136
1137	vhost_hlen = nvq->vhost_hlen;
1138	sock_hlen = nvq->sock_hlen;
1139
1140	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1141		vq->log : NULL;
1142	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1143
1144	do {
1145		sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1146						      &busyloop_intr);
1147		if (!sock_len)
1148			break;
1149		sock_len += sock_hlen;
1150		vhost_len = sock_len + vhost_hlen;
1151		headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1152					vhost_len, &in, vq_log, &log,
1153					likely(mergeable) ? UIO_MAXIOV : 1);
1154		/* On error, stop handling until the next kick. */
1155		if (unlikely(headcount < 0))
1156			goto out;
1157		/* OK, now we need to know about added descriptors. */
1158		if (!headcount) {
1159			if (unlikely(busyloop_intr)) {
1160				vhost_poll_queue(&vq->poll);
1161			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1162				/* They have slipped one in as we were
1163				 * doing that: check again. */
1164				vhost_disable_notify(&net->dev, vq);
1165				continue;
1166			}
1167			/* Nothing new?  Wait for eventfd to tell us
1168			 * they refilled. */
1169			goto out;
1170		}
1171		busyloop_intr = false;
1172		if (nvq->rx_ring)
1173			msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1174		/* On overrun, truncate and discard */
1175		if (unlikely(headcount > UIO_MAXIOV)) {
1176			iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1177			err = sock->ops->recvmsg(sock, &msg,
1178						 1, MSG_DONTWAIT | MSG_TRUNC);
1179			pr_debug("Discarded rx packet: len %zd\n", sock_len);
1180			continue;
1181		}
1182		/* We don't need to be notified again. */
1183		iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1184		fixup = msg.msg_iter;
1185		if (unlikely((vhost_hlen))) {
1186			/* We will supply the header ourselves
1187			 * TODO: support TSO.
1188			 */
1189			iov_iter_advance(&msg.msg_iter, vhost_hlen);
1190		}
1191		err = sock->ops->recvmsg(sock, &msg,
1192					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1193		/* Userspace might have consumed the packet meanwhile:
1194		 * it's not supposed to do this usually, but might be hard
1195		 * to prevent. Discard data we got (if any) and keep going. */
1196		if (unlikely(err != sock_len)) {
1197			pr_debug("Discarded rx packet: "
1198				 " len %d, expected %zd\n", err, sock_len);
1199			vhost_discard_vq_desc(vq, headcount);
1200			continue;
1201		}
1202		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1203		if (unlikely(vhost_hlen)) {
1204			if (copy_to_iter(&hdr, sizeof(hdr),
1205					 &fixup) != sizeof(hdr)) {
1206				vq_err(vq, "Unable to write vnet_hdr "
1207				       "at addr %p\n", vq->iov->iov_base);
1208				goto out;
1209			}
1210		} else {
1211			/* Header came from socket; we'll need to patch
1212			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1213			 */
1214			iov_iter_advance(&fixup, sizeof(hdr));
1215		}
1216		/* TODO: Should check and handle checksum. */
1217
1218		num_buffers = cpu_to_vhost16(vq, headcount);
1219		if (likely(mergeable) &&
1220		    copy_to_iter(&num_buffers, sizeof num_buffers,
1221				 &fixup) != sizeof num_buffers) {
1222			vq_err(vq, "Failed num_buffers write");
1223			vhost_discard_vq_desc(vq, headcount);
1224			goto out;
1225		}
1226		nvq->done_idx += headcount;
1227		if (nvq->done_idx > VHOST_NET_BATCH)
1228			vhost_net_signal_used(nvq);
1229		if (unlikely(vq_log))
1230			vhost_log_write(vq, vq_log, log, vhost_len,
1231					vq->iov, in);
1232		total_len += vhost_len;
1233	} while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1234
1235	if (unlikely(busyloop_intr))
1236		vhost_poll_queue(&vq->poll);
1237	else if (!sock_len)
1238		vhost_net_enable_vq(net, vq);
1239out:
1240	vhost_net_signal_used(nvq);
1241	mutex_unlock(&vq->mutex);
1242}
1243
1244static void handle_tx_kick(struct vhost_work *work)
1245{
1246	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1247						  poll.work);
1248	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1249
1250	handle_tx(net);
1251}
1252
1253static void handle_rx_kick(struct vhost_work *work)
1254{
1255	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1256						  poll.work);
1257	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1258
1259	handle_rx(net);
1260}
1261
1262static void handle_tx_net(struct vhost_work *work)
1263{
1264	struct vhost_net *net = container_of(work, struct vhost_net,
1265					     poll[VHOST_NET_VQ_TX].work);
1266	handle_tx(net);
1267}
1268
1269static void handle_rx_net(struct vhost_work *work)
1270{
1271	struct vhost_net *net = container_of(work, struct vhost_net,
1272					     poll[VHOST_NET_VQ_RX].work);
1273	handle_rx(net);
1274}
1275
1276static int vhost_net_open(struct inode *inode, struct file *f)
1277{
1278	struct vhost_net *n;
1279	struct vhost_dev *dev;
1280	struct vhost_virtqueue **vqs;
1281	void **queue;
1282	struct xdp_buff *xdp;
1283	int i;
1284
1285	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1286	if (!n)
1287		return -ENOMEM;
1288	vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1289	if (!vqs) {
1290		kvfree(n);
1291		return -ENOMEM;
1292	}
1293
1294	queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1295			      GFP_KERNEL);
1296	if (!queue) {
1297		kfree(vqs);
1298		kvfree(n);
1299		return -ENOMEM;
1300	}
1301	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1302
1303	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1304	if (!xdp) {
1305		kfree(vqs);
1306		kvfree(n);
1307		kfree(queue);
1308		return -ENOMEM;
1309	}
1310	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1311
1312	dev = &n->dev;
1313	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1314	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1315	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1316	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1317	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1318		n->vqs[i].ubufs = NULL;
1319		n->vqs[i].ubuf_info = NULL;
1320		n->vqs[i].upend_idx = 0;
1321		n->vqs[i].done_idx = 0;
1322		n->vqs[i].batched_xdp = 0;
1323		n->vqs[i].vhost_hlen = 0;
1324		n->vqs[i].sock_hlen = 0;
1325		n->vqs[i].rx_ring = NULL;
1326		vhost_net_buf_init(&n->vqs[i].rxq);
1327	}
1328	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1329		       UIO_MAXIOV + VHOST_NET_BATCH,
1330		       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1331		       NULL);
1332
1333	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1334	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1335
1336	f->private_data = n;
1337	n->page_frag.page = NULL;
1338	n->refcnt_bias = 0;
1339
1340	return 0;
1341}
1342
1343static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1344					struct vhost_virtqueue *vq)
1345{
1346	struct socket *sock;
1347	struct vhost_net_virtqueue *nvq =
1348		container_of(vq, struct vhost_net_virtqueue, vq);
1349
1350	mutex_lock(&vq->mutex);
1351	sock = vhost_vq_get_backend(vq);
1352	vhost_net_disable_vq(n, vq);
1353	vhost_vq_set_backend(vq, NULL);
1354	vhost_net_buf_unproduce(nvq);
1355	nvq->rx_ring = NULL;
1356	mutex_unlock(&vq->mutex);
1357	return sock;
1358}
1359
1360static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1361			   struct socket **rx_sock)
1362{
1363	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1364	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1365}
1366
1367static void vhost_net_flush_vq(struct vhost_net *n, int index)
1368{
1369	vhost_poll_flush(n->poll + index);
1370	vhost_poll_flush(&n->vqs[index].vq.poll);
1371}
1372
1373static void vhost_net_flush(struct vhost_net *n)
1374{
1375	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1376	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1377	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1378		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1379		n->tx_flush = true;
1380		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1381		/* Wait for all lower device DMAs done. */
1382		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1383		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1384		n->tx_flush = false;
1385		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1386		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1387	}
1388}
1389
1390static int vhost_net_release(struct inode *inode, struct file *f)
1391{
1392	struct vhost_net *n = f->private_data;
1393	struct socket *tx_sock;
1394	struct socket *rx_sock;
1395
1396	vhost_net_stop(n, &tx_sock, &rx_sock);
1397	vhost_net_flush(n);
1398	vhost_dev_stop(&n->dev);
1399	vhost_dev_cleanup(&n->dev);
1400	vhost_net_vq_reset(n);
1401	if (tx_sock)
1402		sockfd_put(tx_sock);
1403	if (rx_sock)
1404		sockfd_put(rx_sock);
1405	/* Make sure no callbacks are outstanding */
1406	synchronize_rcu();
1407	/* We do an extra flush before freeing memory,
1408	 * since jobs can re-queue themselves. */
1409	vhost_net_flush(n);
1410	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1411	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1412	kfree(n->dev.vqs);
1413	if (n->page_frag.page)
1414		__page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1415	kvfree(n);
1416	return 0;
1417}
1418
1419static struct socket *get_raw_socket(int fd)
1420{
 
 
 
 
1421	int r;
1422	struct socket *sock = sockfd_lookup(fd, &r);
1423
1424	if (!sock)
1425		return ERR_PTR(-ENOTSOCK);
1426
1427	/* Parameter checking */
1428	if (sock->sk->sk_type != SOCK_RAW) {
1429		r = -ESOCKTNOSUPPORT;
1430		goto err;
1431	}
1432
1433	if (sock->sk->sk_family != AF_PACKET) {
 
 
 
 
1434		r = -EPFNOSUPPORT;
1435		goto err;
1436	}
1437	return sock;
1438err:
1439	sockfd_put(sock);
1440	return ERR_PTR(r);
1441}
1442
1443static struct ptr_ring *get_tap_ptr_ring(int fd)
1444{
1445	struct ptr_ring *ring;
1446	struct file *file = fget(fd);
1447
1448	if (!file)
1449		return NULL;
1450	ring = tun_get_tx_ring(file);
1451	if (!IS_ERR(ring))
1452		goto out;
1453	ring = tap_get_ptr_ring(file);
1454	if (!IS_ERR(ring))
1455		goto out;
1456	ring = NULL;
1457out:
1458	fput(file);
1459	return ring;
1460}
1461
1462static struct socket *get_tap_socket(int fd)
1463{
1464	struct file *file = fget(fd);
1465	struct socket *sock;
1466
1467	if (!file)
1468		return ERR_PTR(-EBADF);
1469	sock = tun_get_socket(file);
1470	if (!IS_ERR(sock))
1471		return sock;
1472	sock = tap_get_socket(file);
1473	if (IS_ERR(sock))
1474		fput(file);
1475	return sock;
1476}
1477
1478static struct socket *get_socket(int fd)
1479{
1480	struct socket *sock;
1481
1482	/* special case to disable backend */
1483	if (fd == -1)
1484		return NULL;
1485	sock = get_raw_socket(fd);
1486	if (!IS_ERR(sock))
1487		return sock;
1488	sock = get_tap_socket(fd);
1489	if (!IS_ERR(sock))
1490		return sock;
1491	return ERR_PTR(-ENOTSOCK);
1492}
1493
1494static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1495{
1496	struct socket *sock, *oldsock;
1497	struct vhost_virtqueue *vq;
1498	struct vhost_net_virtqueue *nvq;
1499	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1500	int r;
1501
1502	mutex_lock(&n->dev.mutex);
1503	r = vhost_dev_check_owner(&n->dev);
1504	if (r)
1505		goto err;
1506
1507	if (index >= VHOST_NET_VQ_MAX) {
1508		r = -ENOBUFS;
1509		goto err;
1510	}
1511	vq = &n->vqs[index].vq;
1512	nvq = &n->vqs[index];
1513	mutex_lock(&vq->mutex);
1514
1515	/* Verify that ring has been setup correctly. */
1516	if (!vhost_vq_access_ok(vq)) {
1517		r = -EFAULT;
1518		goto err_vq;
1519	}
1520	sock = get_socket(fd);
1521	if (IS_ERR(sock)) {
1522		r = PTR_ERR(sock);
1523		goto err_vq;
1524	}
1525
1526	/* start polling new socket */
1527	oldsock = vhost_vq_get_backend(vq);
1528	if (sock != oldsock) {
1529		ubufs = vhost_net_ubuf_alloc(vq,
1530					     sock && vhost_sock_zcopy(sock));
1531		if (IS_ERR(ubufs)) {
1532			r = PTR_ERR(ubufs);
1533			goto err_ubufs;
1534		}
1535
1536		vhost_net_disable_vq(n, vq);
1537		vhost_vq_set_backend(vq, sock);
1538		vhost_net_buf_unproduce(nvq);
1539		r = vhost_vq_init_access(vq);
1540		if (r)
1541			goto err_used;
1542		r = vhost_net_enable_vq(n, vq);
1543		if (r)
1544			goto err_used;
1545		if (index == VHOST_NET_VQ_RX)
1546			nvq->rx_ring = get_tap_ptr_ring(fd);
1547
1548		oldubufs = nvq->ubufs;
1549		nvq->ubufs = ubufs;
1550
1551		n->tx_packets = 0;
1552		n->tx_zcopy_err = 0;
1553		n->tx_flush = false;
1554	}
1555
1556	mutex_unlock(&vq->mutex);
1557
1558	if (oldubufs) {
1559		vhost_net_ubuf_put_wait_and_free(oldubufs);
1560		mutex_lock(&vq->mutex);
1561		vhost_zerocopy_signal_used(n, vq);
1562		mutex_unlock(&vq->mutex);
1563	}
1564
1565	if (oldsock) {
1566		vhost_net_flush_vq(n, index);
1567		sockfd_put(oldsock);
1568	}
1569
1570	mutex_unlock(&n->dev.mutex);
1571	return 0;
1572
1573err_used:
1574	vhost_vq_set_backend(vq, oldsock);
1575	vhost_net_enable_vq(n, vq);
1576	if (ubufs)
1577		vhost_net_ubuf_put_wait_and_free(ubufs);
1578err_ubufs:
1579	if (sock)
1580		sockfd_put(sock);
1581err_vq:
1582	mutex_unlock(&vq->mutex);
1583err:
1584	mutex_unlock(&n->dev.mutex);
1585	return r;
1586}
1587
1588static long vhost_net_reset_owner(struct vhost_net *n)
1589{
1590	struct socket *tx_sock = NULL;
1591	struct socket *rx_sock = NULL;
1592	long err;
1593	struct vhost_iotlb *umem;
1594
1595	mutex_lock(&n->dev.mutex);
1596	err = vhost_dev_check_owner(&n->dev);
1597	if (err)
1598		goto done;
1599	umem = vhost_dev_reset_owner_prepare();
1600	if (!umem) {
1601		err = -ENOMEM;
1602		goto done;
1603	}
1604	vhost_net_stop(n, &tx_sock, &rx_sock);
1605	vhost_net_flush(n);
1606	vhost_dev_stop(&n->dev);
1607	vhost_dev_reset_owner(&n->dev, umem);
1608	vhost_net_vq_reset(n);
1609done:
1610	mutex_unlock(&n->dev.mutex);
1611	if (tx_sock)
1612		sockfd_put(tx_sock);
1613	if (rx_sock)
1614		sockfd_put(rx_sock);
1615	return err;
1616}
1617
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1618static int vhost_net_set_features(struct vhost_net *n, u64 features)
1619{
1620	size_t vhost_hlen, sock_hlen, hdr_len;
1621	int i;
1622
1623	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1624			       (1ULL << VIRTIO_F_VERSION_1))) ?
1625			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1626			sizeof(struct virtio_net_hdr);
1627	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1628		/* vhost provides vnet_hdr */
1629		vhost_hlen = hdr_len;
1630		sock_hlen = 0;
1631	} else {
1632		/* socket provides vnet_hdr */
1633		vhost_hlen = 0;
1634		sock_hlen = hdr_len;
1635	}
1636	mutex_lock(&n->dev.mutex);
1637	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1638	    !vhost_log_access_ok(&n->dev))
1639		goto out_unlock;
1640
1641	if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1642		if (vhost_init_device_iotlb(&n->dev, true))
1643			goto out_unlock;
1644	}
1645
1646	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1647		mutex_lock(&n->vqs[i].vq.mutex);
1648		n->vqs[i].vq.acked_features = features;
1649		n->vqs[i].vhost_hlen = vhost_hlen;
1650		n->vqs[i].sock_hlen = sock_hlen;
1651		mutex_unlock(&n->vqs[i].vq.mutex);
1652	}
1653	mutex_unlock(&n->dev.mutex);
1654	return 0;
1655
1656out_unlock:
1657	mutex_unlock(&n->dev.mutex);
1658	return -EFAULT;
1659}
1660
1661static long vhost_net_set_owner(struct vhost_net *n)
1662{
1663	int r;
1664
1665	mutex_lock(&n->dev.mutex);
1666	if (vhost_dev_has_owner(&n->dev)) {
1667		r = -EBUSY;
1668		goto out;
1669	}
1670	r = vhost_net_set_ubuf_info(n);
1671	if (r)
1672		goto out;
1673	r = vhost_dev_set_owner(&n->dev);
1674	if (r)
1675		vhost_net_clear_ubuf_info(n);
1676	vhost_net_flush(n);
1677out:
1678	mutex_unlock(&n->dev.mutex);
1679	return r;
1680}
1681
1682static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1683			    unsigned long arg)
1684{
1685	struct vhost_net *n = f->private_data;
1686	void __user *argp = (void __user *)arg;
1687	u64 __user *featurep = argp;
1688	struct vhost_vring_file backend;
1689	u64 features;
1690	int r;
1691
1692	switch (ioctl) {
1693	case VHOST_NET_SET_BACKEND:
1694		if (copy_from_user(&backend, argp, sizeof backend))
1695			return -EFAULT;
1696		return vhost_net_set_backend(n, backend.index, backend.fd);
1697	case VHOST_GET_FEATURES:
1698		features = VHOST_NET_FEATURES;
1699		if (copy_to_user(featurep, &features, sizeof features))
1700			return -EFAULT;
1701		return 0;
1702	case VHOST_SET_FEATURES:
1703		if (copy_from_user(&features, featurep, sizeof features))
1704			return -EFAULT;
1705		if (features & ~VHOST_NET_FEATURES)
1706			return -EOPNOTSUPP;
1707		return vhost_net_set_features(n, features);
1708	case VHOST_GET_BACKEND_FEATURES:
1709		features = VHOST_NET_BACKEND_FEATURES;
1710		if (copy_to_user(featurep, &features, sizeof(features)))
1711			return -EFAULT;
1712		return 0;
1713	case VHOST_SET_BACKEND_FEATURES:
1714		if (copy_from_user(&features, featurep, sizeof(features)))
1715			return -EFAULT;
1716		if (features & ~VHOST_NET_BACKEND_FEATURES)
1717			return -EOPNOTSUPP;
1718		vhost_set_backend_features(&n->dev, features);
1719		return 0;
1720	case VHOST_RESET_OWNER:
1721		return vhost_net_reset_owner(n);
1722	case VHOST_SET_OWNER:
1723		return vhost_net_set_owner(n);
1724	default:
1725		mutex_lock(&n->dev.mutex);
1726		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1727		if (r == -ENOIOCTLCMD)
1728			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1729		else
1730			vhost_net_flush(n);
1731		mutex_unlock(&n->dev.mutex);
1732		return r;
1733	}
1734}
1735
 
 
 
 
 
 
 
 
1736static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1737{
1738	struct file *file = iocb->ki_filp;
1739	struct vhost_net *n = file->private_data;
1740	struct vhost_dev *dev = &n->dev;
1741	int noblock = file->f_flags & O_NONBLOCK;
1742
1743	return vhost_chr_read_iter(dev, to, noblock);
1744}
1745
1746static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1747					struct iov_iter *from)
1748{
1749	struct file *file = iocb->ki_filp;
1750	struct vhost_net *n = file->private_data;
1751	struct vhost_dev *dev = &n->dev;
1752
1753	return vhost_chr_write_iter(dev, from);
1754}
1755
1756static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1757{
1758	struct vhost_net *n = file->private_data;
1759	struct vhost_dev *dev = &n->dev;
1760
1761	return vhost_chr_poll(file, dev, wait);
1762}
1763
1764static const struct file_operations vhost_net_fops = {
1765	.owner          = THIS_MODULE,
1766	.release        = vhost_net_release,
1767	.read_iter      = vhost_net_chr_read_iter,
1768	.write_iter     = vhost_net_chr_write_iter,
1769	.poll           = vhost_net_chr_poll,
1770	.unlocked_ioctl = vhost_net_ioctl,
1771	.compat_ioctl   = compat_ptr_ioctl,
 
 
1772	.open           = vhost_net_open,
1773	.llseek		= noop_llseek,
1774};
1775
1776static struct miscdevice vhost_net_misc = {
1777	.minor = VHOST_NET_MINOR,
1778	.name = "vhost-net",
1779	.fops = &vhost_net_fops,
1780};
1781
1782static int vhost_net_init(void)
1783{
1784	if (experimental_zcopytx)
1785		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1786	return misc_register(&vhost_net_misc);
1787}
1788module_init(vhost_net_init);
1789
1790static void vhost_net_exit(void)
1791{
1792	misc_deregister(&vhost_net_misc);
1793}
1794module_exit(vhost_net_exit);
1795
1796MODULE_VERSION("0.0.1");
1797MODULE_LICENSE("GPL v2");
1798MODULE_AUTHOR("Michael S. Tsirkin");
1799MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1800MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1801MODULE_ALIAS("devname:vhost-net");