xref: /kernel/linux/linux-5.10/drivers/vhost/net.c (revision 8c2ecf20)
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 i, err;
470
471	if (nvq->batched_xdp == 0)
472		goto signal_used;
473
474	msghdr->msg_control = &ctl;
475	msghdr->msg_controllen = sizeof(ctl);
476	err = sock->ops->sendmsg(sock, msghdr, 0);
477	if (unlikely(err < 0)) {
478		vq_err(&nvq->vq, "Fail to batch sending packets\n");
479
480		/* free pages owned by XDP; since this is an unlikely error path,
481		 * keep it simple and avoid more complex bulk update for the
482		 * used pages
483		 */
484		for (i = 0; i < nvq->batched_xdp; ++i)
485			put_page(virt_to_head_page(nvq->xdp[i].data));
486		nvq->batched_xdp = 0;
487		nvq->done_idx = 0;
488		return;
489	}
490
491signal_used:
492	vhost_net_signal_used(nvq);
493	nvq->batched_xdp = 0;
494}
495
496static int sock_has_rx_data(struct socket *sock)
497{
498	if (unlikely(!sock))
499		return 0;
500
501	if (sock->ops->peek_len)
502		return sock->ops->peek_len(sock);
503
504	return skb_queue_empty(&sock->sk->sk_receive_queue);
505}
506
507static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
508					  struct vhost_virtqueue *vq)
509{
510	if (!vhost_vq_avail_empty(&net->dev, vq)) {
511		vhost_poll_queue(&vq->poll);
512	} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
513		vhost_disable_notify(&net->dev, vq);
514		vhost_poll_queue(&vq->poll);
515	}
516}
517
518static void vhost_net_busy_poll(struct vhost_net *net,
519				struct vhost_virtqueue *rvq,
520				struct vhost_virtqueue *tvq,
521				bool *busyloop_intr,
522				bool poll_rx)
523{
524	unsigned long busyloop_timeout;
525	unsigned long endtime;
526	struct socket *sock;
527	struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
528
529	/* Try to hold the vq mutex of the paired virtqueue. We can't
530	 * use mutex_lock() here since we could not guarantee a
531	 * consistenet lock ordering.
532	 */
533	if (!mutex_trylock(&vq->mutex))
534		return;
535
536	vhost_disable_notify(&net->dev, vq);
537	sock = vhost_vq_get_backend(rvq);
538
539	busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
540				     tvq->busyloop_timeout;
541
542	preempt_disable();
543	endtime = busy_clock() + busyloop_timeout;
544
545	while (vhost_can_busy_poll(endtime)) {
546		if (vhost_has_work(&net->dev)) {
547			*busyloop_intr = true;
548			break;
549		}
550
551		if ((sock_has_rx_data(sock) &&
552		     !vhost_vq_avail_empty(&net->dev, rvq)) ||
553		    !vhost_vq_avail_empty(&net->dev, tvq))
554			break;
555
556		cpu_relax();
557	}
558
559	preempt_enable();
560
561	if (poll_rx || sock_has_rx_data(sock))
562		vhost_net_busy_poll_try_queue(net, vq);
563	else if (!poll_rx) /* On tx here, sock has no rx data. */
564		vhost_enable_notify(&net->dev, rvq);
565
566	mutex_unlock(&vq->mutex);
567}
568
569static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
570				    struct vhost_net_virtqueue *tnvq,
571				    unsigned int *out_num, unsigned int *in_num,
572				    struct msghdr *msghdr, bool *busyloop_intr)
573{
574	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
575	struct vhost_virtqueue *rvq = &rnvq->vq;
576	struct vhost_virtqueue *tvq = &tnvq->vq;
577
578	int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
579				  out_num, in_num, NULL, NULL);
580
581	if (r == tvq->num && tvq->busyloop_timeout) {
582		/* Flush batched packets first */
583		if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
584			vhost_tx_batch(net, tnvq,
585				       vhost_vq_get_backend(tvq),
586				       msghdr);
587
588		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
589
590		r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
591				      out_num, in_num, NULL, NULL);
592	}
593
594	return r;
595}
596
597static bool vhost_exceeds_maxpend(struct vhost_net *net)
598{
599	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
600	struct vhost_virtqueue *vq = &nvq->vq;
601
602	return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
603	       min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
604}
605
606static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
607			    size_t hdr_size, int out)
608{
609	/* Skip header. TODO: support TSO. */
610	size_t len = iov_length(vq->iov, out);
611
612	iov_iter_init(iter, WRITE, vq->iov, out, len);
613	iov_iter_advance(iter, hdr_size);
614
615	return iov_iter_count(iter);
616}
617
618static int get_tx_bufs(struct vhost_net *net,
619		       struct vhost_net_virtqueue *nvq,
620		       struct msghdr *msg,
621		       unsigned int *out, unsigned int *in,
622		       size_t *len, bool *busyloop_intr)
623{
624	struct vhost_virtqueue *vq = &nvq->vq;
625	int ret;
626
627	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
628
629	if (ret < 0 || ret == vq->num)
630		return ret;
631
632	if (*in) {
633		vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
634			*out, *in);
635		return -EFAULT;
636	}
637
638	/* Sanity check */
639	*len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
640	if (*len == 0) {
641		vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
642			*len, nvq->vhost_hlen);
643		return -EFAULT;
644	}
645
646	return ret;
647}
648
649static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
650{
651	return total_len < VHOST_NET_WEIGHT &&
652	       !vhost_vq_avail_empty(vq->dev, vq);
653}
654
655static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
656				       struct page_frag *pfrag, gfp_t gfp)
657{
658	if (pfrag->page) {
659		if (pfrag->offset + sz <= pfrag->size)
660			return true;
661		__page_frag_cache_drain(pfrag->page, net->refcnt_bias);
662	}
663
664	pfrag->offset = 0;
665	net->refcnt_bias = 0;
666	if (SKB_FRAG_PAGE_ORDER) {
667		/* Avoid direct reclaim but allow kswapd to wake */
668		pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
669					  __GFP_COMP | __GFP_NOWARN |
670					  __GFP_NORETRY,
671					  SKB_FRAG_PAGE_ORDER);
672		if (likely(pfrag->page)) {
673			pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
674			goto done;
675		}
676	}
677	pfrag->page = alloc_page(gfp);
678	if (likely(pfrag->page)) {
679		pfrag->size = PAGE_SIZE;
680		goto done;
681	}
682	return false;
683
684done:
685	net->refcnt_bias = USHRT_MAX;
686	page_ref_add(pfrag->page, USHRT_MAX - 1);
687	return true;
688}
689
690#define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
691
692static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
693			       struct iov_iter *from)
694{
695	struct vhost_virtqueue *vq = &nvq->vq;
696	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
697					     dev);
698	struct socket *sock = vhost_vq_get_backend(vq);
699	struct page_frag *alloc_frag = &net->page_frag;
700	struct virtio_net_hdr *gso;
701	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
702	struct tun_xdp_hdr *hdr;
703	size_t len = iov_iter_count(from);
704	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
705	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
706	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
707	int sock_hlen = nvq->sock_hlen;
708	void *buf;
709	int copied;
710
711	if (unlikely(len < nvq->sock_hlen))
712		return -EFAULT;
713
714	if (SKB_DATA_ALIGN(len + pad) +
715	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
716		return -ENOSPC;
717
718	buflen += SKB_DATA_ALIGN(len + pad);
719	alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
720	if (unlikely(!vhost_net_page_frag_refill(net, buflen,
721						 alloc_frag, GFP_KERNEL)))
722		return -ENOMEM;
723
724	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
725	copied = copy_page_from_iter(alloc_frag->page,
726				     alloc_frag->offset +
727				     offsetof(struct tun_xdp_hdr, gso),
728				     sock_hlen, from);
729	if (copied != sock_hlen)
730		return -EFAULT;
731
732	hdr = buf;
733	gso = &hdr->gso;
734
735	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
736	    vhost16_to_cpu(vq, gso->csum_start) +
737	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
738	    vhost16_to_cpu(vq, gso->hdr_len)) {
739		gso->hdr_len = cpu_to_vhost16(vq,
740			       vhost16_to_cpu(vq, gso->csum_start) +
741			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
742
743		if (vhost16_to_cpu(vq, gso->hdr_len) > len)
744			return -EINVAL;
745	}
746
747	len -= sock_hlen;
748	copied = copy_page_from_iter(alloc_frag->page,
749				     alloc_frag->offset + pad,
750				     len, from);
751	if (copied != len)
752		return -EFAULT;
753
754	xdp->data_hard_start = buf;
755	xdp->data = buf + pad;
756	xdp->data_end = xdp->data + len;
757	hdr->buflen = buflen;
758	xdp->frame_sz = buflen;
759
760	--net->refcnt_bias;
761	alloc_frag->offset += buflen;
762
763	++nvq->batched_xdp;
764
765	return 0;
766}
767
768static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
769{
770	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
771	struct vhost_virtqueue *vq = &nvq->vq;
772	unsigned out, in;
773	int head;
774	struct msghdr msg = {
775		.msg_name = NULL,
776		.msg_namelen = 0,
777		.msg_control = NULL,
778		.msg_controllen = 0,
779		.msg_flags = MSG_DONTWAIT,
780	};
781	size_t len, total_len = 0;
782	int err;
783	int sent_pkts = 0;
784	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
785
786	do {
787		bool busyloop_intr = false;
788
789		if (nvq->done_idx == VHOST_NET_BATCH)
790			vhost_tx_batch(net, nvq, sock, &msg);
791
792		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
793				   &busyloop_intr);
794		/* On error, stop handling until the next kick. */
795		if (unlikely(head < 0))
796			break;
797		/* Nothing new?  Wait for eventfd to tell us they refilled. */
798		if (head == vq->num) {
799			if (unlikely(busyloop_intr)) {
800				vhost_poll_queue(&vq->poll);
801			} else if (unlikely(vhost_enable_notify(&net->dev,
802								vq))) {
803				vhost_disable_notify(&net->dev, vq);
804				continue;
805			}
806			break;
807		}
808
809		total_len += len;
810
811		/* For simplicity, TX batching is only enabled if
812		 * sndbuf is unlimited.
813		 */
814		if (sock_can_batch) {
815			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
816			if (!err) {
817				goto done;
818			} else if (unlikely(err != -ENOSPC)) {
819				vhost_tx_batch(net, nvq, sock, &msg);
820				vhost_discard_vq_desc(vq, 1);
821				vhost_net_enable_vq(net, vq);
822				break;
823			}
824
825			/* We can't build XDP buff, go for single
826			 * packet path but let's flush batched
827			 * packets.
828			 */
829			vhost_tx_batch(net, nvq, sock, &msg);
830			msg.msg_control = NULL;
831		} else {
832			if (tx_can_batch(vq, total_len))
833				msg.msg_flags |= MSG_MORE;
834			else
835				msg.msg_flags &= ~MSG_MORE;
836		}
837
838		/* TODO: Check specific error and bomb out unless ENOBUFS? */
839		err = sock->ops->sendmsg(sock, &msg, len);
840		if (unlikely(err < 0)) {
841			vhost_discard_vq_desc(vq, 1);
842			vhost_net_enable_vq(net, vq);
843			break;
844		}
845		if (err != len)
846			pr_debug("Truncated TX packet: len %d != %zd\n",
847				 err, len);
848done:
849		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
850		vq->heads[nvq->done_idx].len = 0;
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 *ubuf;
875	bool zcopy_used;
876	int sent_pkts = 0;
877
878	do {
879		bool busyloop_intr;
880
881		/* Release DMAs done buffers first */
882		vhost_zerocopy_signal_used(net, vq);
883
884		busyloop_intr = false;
885		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
886				   &busyloop_intr);
887		/* On error, stop handling until the next kick. */
888		if (unlikely(head < 0))
889			break;
890		/* Nothing new?  Wait for eventfd to tell us they refilled. */
891		if (head == vq->num) {
892			if (unlikely(busyloop_intr)) {
893				vhost_poll_queue(&vq->poll);
894			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
895				vhost_disable_notify(&net->dev, vq);
896				continue;
897			}
898			break;
899		}
900
901		zcopy_used = len >= VHOST_GOODCOPY_LEN
902			     && !vhost_exceeds_maxpend(net)
903			     && vhost_net_tx_select_zcopy(net);
904
905		/* use msg_control to pass vhost zerocopy ubuf info to skb */
906		if (zcopy_used) {
907			ubuf = nvq->ubuf_info + nvq->upend_idx;
908			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
909			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
910			ubuf->callback = vhost_zerocopy_callback;
911			ubuf->ctx = nvq->ubufs;
912			ubuf->desc = nvq->upend_idx;
913			refcount_set(&ubuf->refcnt, 1);
914			msg.msg_control = &ctl;
915			ctl.type = TUN_MSG_UBUF;
916			ctl.ptr = ubuf;
917			msg.msg_controllen = sizeof(ctl);
918			ubufs = nvq->ubufs;
919			atomic_inc(&ubufs->refcount);
920			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
921		} else {
922			msg.msg_control = NULL;
923			ubufs = NULL;
924		}
925		total_len += len;
926		if (tx_can_batch(vq, total_len) &&
927		    likely(!vhost_exceeds_maxpend(net))) {
928			msg.msg_flags |= MSG_MORE;
929		} else {
930			msg.msg_flags &= ~MSG_MORE;
931		}
932
933		/* TODO: Check specific error and bomb out unless ENOBUFS? */
934		err = sock->ops->sendmsg(sock, &msg, len);
935		if (unlikely(err < 0)) {
936			if (zcopy_used) {
937				if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
938					vhost_net_ubuf_put(ubufs);
939				nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
940					% UIO_MAXIOV;
941			}
942			vhost_discard_vq_desc(vq, 1);
943			vhost_net_enable_vq(net, vq);
944			break;
945		}
946		if (err != len)
947			pr_debug("Truncated TX packet: "
948				 " len %d != %zd\n", err, len);
949		if (!zcopy_used)
950			vhost_add_used_and_signal(&net->dev, vq, head, 0);
951		else
952			vhost_zerocopy_signal_used(net, vq);
953		vhost_net_tx_packet(net);
954	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
955}
956
957/* Expects to be always run from workqueue - which acts as
958 * read-size critical section for our kind of RCU. */
959static void handle_tx(struct vhost_net *net)
960{
961	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
962	struct vhost_virtqueue *vq = &nvq->vq;
963	struct socket *sock;
964
965	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
966	sock = vhost_vq_get_backend(vq);
967	if (!sock)
968		goto out;
969
970	if (!vq_meta_prefetch(vq))
971		goto out;
972
973	vhost_disable_notify(&net->dev, vq);
974	vhost_net_disable_vq(net, vq);
975
976	if (vhost_sock_zcopy(sock))
977		handle_tx_zerocopy(net, sock);
978	else
979		handle_tx_copy(net, sock);
980
981out:
982	mutex_unlock(&vq->mutex);
983}
984
985static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
986{
987	struct sk_buff *head;
988	int len = 0;
989	unsigned long flags;
990
991	if (rvq->rx_ring)
992		return vhost_net_buf_peek(rvq);
993
994	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
995	head = skb_peek(&sk->sk_receive_queue);
996	if (likely(head)) {
997		len = head->len;
998		if (skb_vlan_tag_present(head))
999			len += VLAN_HLEN;
1000	}
1001
1002	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1003	return len;
1004}
1005
1006static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1007				      bool *busyloop_intr)
1008{
1009	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1010	struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1011	struct vhost_virtqueue *rvq = &rnvq->vq;
1012	struct vhost_virtqueue *tvq = &tnvq->vq;
1013	int len = peek_head_len(rnvq, sk);
1014
1015	if (!len && rvq->busyloop_timeout) {
1016		/* Flush batched heads first */
1017		vhost_net_signal_used(rnvq);
1018		/* Both tx vq and rx socket were polled here */
1019		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1020
1021		len = peek_head_len(rnvq, sk);
1022	}
1023
1024	return len;
1025}
1026
1027/* This is a multi-buffer version of vhost_get_desc, that works if
1028 *	vq has read descriptors only.
1029 * @vq		- the relevant virtqueue
1030 * @datalen	- data length we'll be reading
1031 * @iovcount	- returned count of io vectors we fill
1032 * @log		- vhost log
1033 * @log_num	- log offset
1034 * @quota       - headcount quota, 1 for big buffer
1035 *	returns number of buffer heads allocated, negative on error
1036 */
1037static int get_rx_bufs(struct vhost_virtqueue *vq,
1038		       struct vring_used_elem *heads,
1039		       int datalen,
1040		       unsigned *iovcount,
1041		       struct vhost_log *log,
1042		       unsigned *log_num,
1043		       unsigned int quota)
1044{
1045	unsigned int out, in;
1046	int seg = 0;
1047	int headcount = 0;
1048	unsigned d;
1049	int r, nlogs = 0;
1050	/* len is always initialized before use since we are always called with
1051	 * datalen > 0.
1052	 */
1053	u32 len;
1054
1055	while (datalen > 0 && headcount < quota) {
1056		if (unlikely(seg >= UIO_MAXIOV)) {
1057			r = -ENOBUFS;
1058			goto err;
1059		}
1060		r = vhost_get_vq_desc(vq, vq->iov + seg,
1061				      ARRAY_SIZE(vq->iov) - seg, &out,
1062				      &in, log, log_num);
1063		if (unlikely(r < 0))
1064			goto err;
1065
1066		d = r;
1067		if (d == vq->num) {
1068			r = 0;
1069			goto err;
1070		}
1071		if (unlikely(out || in <= 0)) {
1072			vq_err(vq, "unexpected descriptor format for RX: "
1073				"out %d, in %d\n", out, in);
1074			r = -EINVAL;
1075			goto err;
1076		}
1077		if (unlikely(log)) {
1078			nlogs += *log_num;
1079			log += *log_num;
1080		}
1081		heads[headcount].id = cpu_to_vhost32(vq, d);
1082		len = iov_length(vq->iov + seg, in);
1083		heads[headcount].len = cpu_to_vhost32(vq, len);
1084		datalen -= len;
1085		++headcount;
1086		seg += in;
1087	}
1088	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1089	*iovcount = seg;
1090	if (unlikely(log))
1091		*log_num = nlogs;
1092
1093	/* Detect overrun */
1094	if (unlikely(datalen > 0)) {
1095		r = UIO_MAXIOV + 1;
1096		goto err;
1097	}
1098	return headcount;
1099err:
1100	vhost_discard_vq_desc(vq, headcount);
1101	return r;
1102}
1103
1104/* Expects to be always run from workqueue - which acts as
1105 * read-size critical section for our kind of RCU. */
1106static void handle_rx(struct vhost_net *net)
1107{
1108	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1109	struct vhost_virtqueue *vq = &nvq->vq;
1110	unsigned in, log;
1111	struct vhost_log *vq_log;
1112	struct msghdr msg = {
1113		.msg_name = NULL,
1114		.msg_namelen = 0,
1115		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
1116		.msg_controllen = 0,
1117		.msg_flags = MSG_DONTWAIT,
1118	};
1119	struct virtio_net_hdr hdr = {
1120		.flags = 0,
1121		.gso_type = VIRTIO_NET_HDR_GSO_NONE
1122	};
1123	size_t total_len = 0;
1124	int err, mergeable;
1125	s16 headcount;
1126	size_t vhost_hlen, sock_hlen;
1127	size_t vhost_len, sock_len;
1128	bool busyloop_intr = false;
1129	struct socket *sock;
1130	struct iov_iter fixup;
1131	__virtio16 num_buffers;
1132	int recv_pkts = 0;
1133
1134	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1135	sock = vhost_vq_get_backend(vq);
1136	if (!sock)
1137		goto out;
1138
1139	if (!vq_meta_prefetch(vq))
1140		goto out;
1141
1142	vhost_disable_notify(&net->dev, vq);
1143	vhost_net_disable_vq(net, vq);
1144
1145	vhost_hlen = nvq->vhost_hlen;
1146	sock_hlen = nvq->sock_hlen;
1147
1148	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1149		vq->log : NULL;
1150	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1151
1152	do {
1153		sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1154						      &busyloop_intr);
1155		if (!sock_len)
1156			break;
1157		sock_len += sock_hlen;
1158		vhost_len = sock_len + vhost_hlen;
1159		headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1160					vhost_len, &in, vq_log, &log,
1161					likely(mergeable) ? UIO_MAXIOV : 1);
1162		/* On error, stop handling until the next kick. */
1163		if (unlikely(headcount < 0))
1164			goto out;
1165		/* OK, now we need to know about added descriptors. */
1166		if (!headcount) {
1167			if (unlikely(busyloop_intr)) {
1168				vhost_poll_queue(&vq->poll);
1169			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1170				/* They have slipped one in as we were
1171				 * doing that: check again. */
1172				vhost_disable_notify(&net->dev, vq);
1173				continue;
1174			}
1175			/* Nothing new?  Wait for eventfd to tell us
1176			 * they refilled. */
1177			goto out;
1178		}
1179		busyloop_intr = false;
1180		if (nvq->rx_ring)
1181			msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1182		/* On overrun, truncate and discard */
1183		if (unlikely(headcount > UIO_MAXIOV)) {
1184			iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1185			err = sock->ops->recvmsg(sock, &msg,
1186						 1, MSG_DONTWAIT | MSG_TRUNC);
1187			pr_debug("Discarded rx packet: len %zd\n", sock_len);
1188			continue;
1189		}
1190		/* We don't need to be notified again. */
1191		iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1192		fixup = msg.msg_iter;
1193		if (unlikely((vhost_hlen))) {
1194			/* We will supply the header ourselves
1195			 * TODO: support TSO.
1196			 */
1197			iov_iter_advance(&msg.msg_iter, vhost_hlen);
1198		}
1199		err = sock->ops->recvmsg(sock, &msg,
1200					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1201		/* Userspace might have consumed the packet meanwhile:
1202		 * it's not supposed to do this usually, but might be hard
1203		 * to prevent. Discard data we got (if any) and keep going. */
1204		if (unlikely(err != sock_len)) {
1205			pr_debug("Discarded rx packet: "
1206				 " len %d, expected %zd\n", err, sock_len);
1207			vhost_discard_vq_desc(vq, headcount);
1208			continue;
1209		}
1210		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1211		if (unlikely(vhost_hlen)) {
1212			if (copy_to_iter(&hdr, sizeof(hdr),
1213					 &fixup) != sizeof(hdr)) {
1214				vq_err(vq, "Unable to write vnet_hdr "
1215				       "at addr %p\n", vq->iov->iov_base);
1216				goto out;
1217			}
1218		} else {
1219			/* Header came from socket; we'll need to patch
1220			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1221			 */
1222			iov_iter_advance(&fixup, sizeof(hdr));
1223		}
1224		/* TODO: Should check and handle checksum. */
1225
1226		num_buffers = cpu_to_vhost16(vq, headcount);
1227		if (likely(mergeable) &&
1228		    copy_to_iter(&num_buffers, sizeof num_buffers,
1229				 &fixup) != sizeof num_buffers) {
1230			vq_err(vq, "Failed num_buffers write");
1231			vhost_discard_vq_desc(vq, headcount);
1232			goto out;
1233		}
1234		nvq->done_idx += headcount;
1235		if (nvq->done_idx > VHOST_NET_BATCH)
1236			vhost_net_signal_used(nvq);
1237		if (unlikely(vq_log))
1238			vhost_log_write(vq, vq_log, log, vhost_len,
1239					vq->iov, in);
1240		total_len += vhost_len;
1241	} while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1242
1243	if (unlikely(busyloop_intr))
1244		vhost_poll_queue(&vq->poll);
1245	else if (!sock_len)
1246		vhost_net_enable_vq(net, vq);
1247out:
1248	vhost_net_signal_used(nvq);
1249	mutex_unlock(&vq->mutex);
1250}
1251
1252static void handle_tx_kick(struct vhost_work *work)
1253{
1254	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1255						  poll.work);
1256	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1257
1258	handle_tx(net);
1259}
1260
1261static void handle_rx_kick(struct vhost_work *work)
1262{
1263	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1264						  poll.work);
1265	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1266
1267	handle_rx(net);
1268}
1269
1270static void handle_tx_net(struct vhost_work *work)
1271{
1272	struct vhost_net *net = container_of(work, struct vhost_net,
1273					     poll[VHOST_NET_VQ_TX].work);
1274	handle_tx(net);
1275}
1276
1277static void handle_rx_net(struct vhost_work *work)
1278{
1279	struct vhost_net *net = container_of(work, struct vhost_net,
1280					     poll[VHOST_NET_VQ_RX].work);
1281	handle_rx(net);
1282}
1283
1284static int vhost_net_open(struct inode *inode, struct file *f)
1285{
1286	struct vhost_net *n;
1287	struct vhost_dev *dev;
1288	struct vhost_virtqueue **vqs;
1289	void **queue;
1290	struct xdp_buff *xdp;
1291	int i;
1292
1293	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1294	if (!n)
1295		return -ENOMEM;
1296	vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1297	if (!vqs) {
1298		kvfree(n);
1299		return -ENOMEM;
1300	}
1301
1302	queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1303			      GFP_KERNEL);
1304	if (!queue) {
1305		kfree(vqs);
1306		kvfree(n);
1307		return -ENOMEM;
1308	}
1309	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1310
1311	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1312	if (!xdp) {
1313		kfree(vqs);
1314		kvfree(n);
1315		kfree(queue);
1316		return -ENOMEM;
1317	}
1318	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1319
1320	dev = &n->dev;
1321	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1322	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1323	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1324	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1325	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1326		n->vqs[i].ubufs = NULL;
1327		n->vqs[i].ubuf_info = NULL;
1328		n->vqs[i].upend_idx = 0;
1329		n->vqs[i].done_idx = 0;
1330		n->vqs[i].batched_xdp = 0;
1331		n->vqs[i].vhost_hlen = 0;
1332		n->vqs[i].sock_hlen = 0;
1333		n->vqs[i].rx_ring = NULL;
1334		vhost_net_buf_init(&n->vqs[i].rxq);
1335	}
1336	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1337		       UIO_MAXIOV + VHOST_NET_BATCH,
1338		       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1339		       NULL);
1340
1341	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1342	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1343
1344	f->private_data = n;
1345	n->page_frag.page = NULL;
1346	n->refcnt_bias = 0;
1347
1348	return 0;
1349}
1350
1351static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1352					struct vhost_virtqueue *vq)
1353{
1354	struct socket *sock;
1355	struct vhost_net_virtqueue *nvq =
1356		container_of(vq, struct vhost_net_virtqueue, vq);
1357
1358	mutex_lock(&vq->mutex);
1359	sock = vhost_vq_get_backend(vq);
1360	vhost_net_disable_vq(n, vq);
1361	vhost_vq_set_backend(vq, NULL);
1362	vhost_net_buf_unproduce(nvq);
1363	nvq->rx_ring = NULL;
1364	mutex_unlock(&vq->mutex);
1365	return sock;
1366}
1367
1368static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1369			   struct socket **rx_sock)
1370{
1371	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1372	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1373}
1374
1375static void vhost_net_flush_vq(struct vhost_net *n, int index)
1376{
1377	vhost_poll_flush(n->poll + index);
1378	vhost_poll_flush(&n->vqs[index].vq.poll);
1379}
1380
1381static void vhost_net_flush(struct vhost_net *n)
1382{
1383	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1384	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1385	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1386		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1387		n->tx_flush = true;
1388		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1389		/* Wait for all lower device DMAs done. */
1390		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1391		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1392		n->tx_flush = false;
1393		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1394		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1395	}
1396}
1397
1398static int vhost_net_release(struct inode *inode, struct file *f)
1399{
1400	struct vhost_net *n = f->private_data;
1401	struct socket *tx_sock;
1402	struct socket *rx_sock;
1403
1404	vhost_net_stop(n, &tx_sock, &rx_sock);
1405	vhost_net_flush(n);
1406	vhost_dev_stop(&n->dev);
1407	vhost_dev_cleanup(&n->dev);
1408	vhost_net_vq_reset(n);
1409	if (tx_sock)
1410		sockfd_put(tx_sock);
1411	if (rx_sock)
1412		sockfd_put(rx_sock);
1413	/* Make sure no callbacks are outstanding */
1414	synchronize_rcu();
1415	/* We do an extra flush before freeing memory,
1416	 * since jobs can re-queue themselves. */
1417	vhost_net_flush(n);
1418	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1419	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1420	kfree(n->dev.vqs);
1421	if (n->page_frag.page)
1422		__page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1423	kvfree(n);
1424	return 0;
1425}
1426
1427static struct socket *get_raw_socket(int fd)
1428{
1429	int r;
1430	struct socket *sock = sockfd_lookup(fd, &r);
1431
1432	if (!sock)
1433		return ERR_PTR(-ENOTSOCK);
1434
1435	/* Parameter checking */
1436	if (sock->sk->sk_type != SOCK_RAW) {
1437		r = -ESOCKTNOSUPPORT;
1438		goto err;
1439	}
1440
1441	if (sock->sk->sk_family != AF_PACKET) {
1442		r = -EPFNOSUPPORT;
1443		goto err;
1444	}
1445	return sock;
1446err:
1447	sockfd_put(sock);
1448	return ERR_PTR(r);
1449}
1450
1451static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1452{
1453	struct ptr_ring *ring;
1454	ring = tun_get_tx_ring(file);
1455	if (!IS_ERR(ring))
1456		goto out;
1457	ring = tap_get_ptr_ring(file);
1458	if (!IS_ERR(ring))
1459		goto out;
1460	ring = NULL;
1461out:
1462	return ring;
1463}
1464
1465static struct socket *get_tap_socket(int fd)
1466{
1467	struct file *file = fget(fd);
1468	struct socket *sock;
1469
1470	if (!file)
1471		return ERR_PTR(-EBADF);
1472	sock = tun_get_socket(file);
1473	if (!IS_ERR(sock))
1474		return sock;
1475	sock = tap_get_socket(file);
1476	if (IS_ERR(sock))
1477		fput(file);
1478	return sock;
1479}
1480
1481static struct socket *get_socket(int fd)
1482{
1483	struct socket *sock;
1484
1485	/* special case to disable backend */
1486	if (fd == -1)
1487		return NULL;
1488	sock = get_raw_socket(fd);
1489	if (!IS_ERR(sock))
1490		return sock;
1491	sock = get_tap_socket(fd);
1492	if (!IS_ERR(sock))
1493		return sock;
1494	return ERR_PTR(-ENOTSOCK);
1495}
1496
1497static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1498{
1499	struct socket *sock, *oldsock;
1500	struct vhost_virtqueue *vq;
1501	struct vhost_net_virtqueue *nvq;
1502	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1503	int r;
1504
1505	mutex_lock(&n->dev.mutex);
1506	r = vhost_dev_check_owner(&n->dev);
1507	if (r)
1508		goto err;
1509
1510	if (index >= VHOST_NET_VQ_MAX) {
1511		r = -ENOBUFS;
1512		goto err;
1513	}
1514	vq = &n->vqs[index].vq;
1515	nvq = &n->vqs[index];
1516	mutex_lock(&vq->mutex);
1517
1518	if (fd == -1)
1519		vhost_clear_msg(&n->dev);
1520
1521	/* Verify that ring has been setup correctly. */
1522	if (!vhost_vq_access_ok(vq)) {
1523		r = -EFAULT;
1524		goto err_vq;
1525	}
1526	sock = get_socket(fd);
1527	if (IS_ERR(sock)) {
1528		r = PTR_ERR(sock);
1529		goto err_vq;
1530	}
1531
1532	/* start polling new socket */
1533	oldsock = vhost_vq_get_backend(vq);
1534	if (sock != oldsock) {
1535		ubufs = vhost_net_ubuf_alloc(vq,
1536					     sock && vhost_sock_zcopy(sock));
1537		if (IS_ERR(ubufs)) {
1538			r = PTR_ERR(ubufs);
1539			goto err_ubufs;
1540		}
1541
1542		vhost_net_disable_vq(n, vq);
1543		vhost_vq_set_backend(vq, sock);
1544		vhost_net_buf_unproduce(nvq);
1545		r = vhost_vq_init_access(vq);
1546		if (r)
1547			goto err_used;
1548		r = vhost_net_enable_vq(n, vq);
1549		if (r)
1550			goto err_used;
1551		if (index == VHOST_NET_VQ_RX) {
1552			if (sock)
1553				nvq->rx_ring = get_tap_ptr_ring(sock->file);
1554			else
1555				nvq->rx_ring = NULL;
1556		}
1557
1558		oldubufs = nvq->ubufs;
1559		nvq->ubufs = ubufs;
1560
1561		n->tx_packets = 0;
1562		n->tx_zcopy_err = 0;
1563		n->tx_flush = false;
1564	}
1565
1566	mutex_unlock(&vq->mutex);
1567
1568	if (oldubufs) {
1569		vhost_net_ubuf_put_wait_and_free(oldubufs);
1570		mutex_lock(&vq->mutex);
1571		vhost_zerocopy_signal_used(n, vq);
1572		mutex_unlock(&vq->mutex);
1573	}
1574
1575	if (oldsock) {
1576		vhost_net_flush_vq(n, index);
1577		sockfd_put(oldsock);
1578	}
1579
1580	mutex_unlock(&n->dev.mutex);
1581	return 0;
1582
1583err_used:
1584	vhost_vq_set_backend(vq, oldsock);
1585	vhost_net_enable_vq(n, vq);
1586	if (ubufs)
1587		vhost_net_ubuf_put_wait_and_free(ubufs);
1588err_ubufs:
1589	if (sock)
1590		sockfd_put(sock);
1591err_vq:
1592	mutex_unlock(&vq->mutex);
1593err:
1594	mutex_unlock(&n->dev.mutex);
1595	return r;
1596}
1597
1598static long vhost_net_reset_owner(struct vhost_net *n)
1599{
1600	struct socket *tx_sock = NULL;
1601	struct socket *rx_sock = NULL;
1602	long err;
1603	struct vhost_iotlb *umem;
1604
1605	mutex_lock(&n->dev.mutex);
1606	err = vhost_dev_check_owner(&n->dev);
1607	if (err)
1608		goto done;
1609	umem = vhost_dev_reset_owner_prepare();
1610	if (!umem) {
1611		err = -ENOMEM;
1612		goto done;
1613	}
1614	vhost_net_stop(n, &tx_sock, &rx_sock);
1615	vhost_net_flush(n);
1616	vhost_dev_stop(&n->dev);
1617	vhost_dev_reset_owner(&n->dev, umem);
1618	vhost_net_vq_reset(n);
1619done:
1620	mutex_unlock(&n->dev.mutex);
1621	if (tx_sock)
1622		sockfd_put(tx_sock);
1623	if (rx_sock)
1624		sockfd_put(rx_sock);
1625	return err;
1626}
1627
1628static int vhost_net_set_features(struct vhost_net *n, u64 features)
1629{
1630	size_t vhost_hlen, sock_hlen, hdr_len;
1631	int i;
1632
1633	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1634			       (1ULL << VIRTIO_F_VERSION_1))) ?
1635			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1636			sizeof(struct virtio_net_hdr);
1637	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1638		/* vhost provides vnet_hdr */
1639		vhost_hlen = hdr_len;
1640		sock_hlen = 0;
1641	} else {
1642		/* socket provides vnet_hdr */
1643		vhost_hlen = 0;
1644		sock_hlen = hdr_len;
1645	}
1646	mutex_lock(&n->dev.mutex);
1647	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1648	    !vhost_log_access_ok(&n->dev))
1649		goto out_unlock;
1650
1651	if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1652		if (vhost_init_device_iotlb(&n->dev, true))
1653			goto out_unlock;
1654	}
1655
1656	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1657		mutex_lock(&n->vqs[i].vq.mutex);
1658		n->vqs[i].vq.acked_features = features;
1659		n->vqs[i].vhost_hlen = vhost_hlen;
1660		n->vqs[i].sock_hlen = sock_hlen;
1661		mutex_unlock(&n->vqs[i].vq.mutex);
1662	}
1663	mutex_unlock(&n->dev.mutex);
1664	return 0;
1665
1666out_unlock:
1667	mutex_unlock(&n->dev.mutex);
1668	return -EFAULT;
1669}
1670
1671static long vhost_net_set_owner(struct vhost_net *n)
1672{
1673	int r;
1674
1675	mutex_lock(&n->dev.mutex);
1676	if (vhost_dev_has_owner(&n->dev)) {
1677		r = -EBUSY;
1678		goto out;
1679	}
1680	r = vhost_net_set_ubuf_info(n);
1681	if (r)
1682		goto out;
1683	r = vhost_dev_set_owner(&n->dev);
1684	if (r)
1685		vhost_net_clear_ubuf_info(n);
1686	vhost_net_flush(n);
1687out:
1688	mutex_unlock(&n->dev.mutex);
1689	return r;
1690}
1691
1692static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1693			    unsigned long arg)
1694{
1695	struct vhost_net *n = f->private_data;
1696	void __user *argp = (void __user *)arg;
1697	u64 __user *featurep = argp;
1698	struct vhost_vring_file backend;
1699	u64 features;
1700	int r;
1701
1702	switch (ioctl) {
1703	case VHOST_NET_SET_BACKEND:
1704		if (copy_from_user(&backend, argp, sizeof backend))
1705			return -EFAULT;
1706		return vhost_net_set_backend(n, backend.index, backend.fd);
1707	case VHOST_GET_FEATURES:
1708		features = VHOST_NET_FEATURES;
1709		if (copy_to_user(featurep, &features, sizeof features))
1710			return -EFAULT;
1711		return 0;
1712	case VHOST_SET_FEATURES:
1713		if (copy_from_user(&features, featurep, sizeof features))
1714			return -EFAULT;
1715		if (features & ~VHOST_NET_FEATURES)
1716			return -EOPNOTSUPP;
1717		return vhost_net_set_features(n, features);
1718	case VHOST_GET_BACKEND_FEATURES:
1719		features = VHOST_NET_BACKEND_FEATURES;
1720		if (copy_to_user(featurep, &features, sizeof(features)))
1721			return -EFAULT;
1722		return 0;
1723	case VHOST_SET_BACKEND_FEATURES:
1724		if (copy_from_user(&features, featurep, sizeof(features)))
1725			return -EFAULT;
1726		if (features & ~VHOST_NET_BACKEND_FEATURES)
1727			return -EOPNOTSUPP;
1728		vhost_set_backend_features(&n->dev, features);
1729		return 0;
1730	case VHOST_RESET_OWNER:
1731		return vhost_net_reset_owner(n);
1732	case VHOST_SET_OWNER:
1733		return vhost_net_set_owner(n);
1734	default:
1735		mutex_lock(&n->dev.mutex);
1736		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1737		if (r == -ENOIOCTLCMD)
1738			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1739		else
1740			vhost_net_flush(n);
1741		mutex_unlock(&n->dev.mutex);
1742		return r;
1743	}
1744}
1745
1746static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1747{
1748	struct file *file = iocb->ki_filp;
1749	struct vhost_net *n = file->private_data;
1750	struct vhost_dev *dev = &n->dev;
1751	int noblock = file->f_flags & O_NONBLOCK;
1752
1753	return vhost_chr_read_iter(dev, to, noblock);
1754}
1755
1756static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1757					struct iov_iter *from)
1758{
1759	struct file *file = iocb->ki_filp;
1760	struct vhost_net *n = file->private_data;
1761	struct vhost_dev *dev = &n->dev;
1762
1763	return vhost_chr_write_iter(dev, from);
1764}
1765
1766static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1767{
1768	struct vhost_net *n = file->private_data;
1769	struct vhost_dev *dev = &n->dev;
1770
1771	return vhost_chr_poll(file, dev, wait);
1772}
1773
1774static const struct file_operations vhost_net_fops = {
1775	.owner          = THIS_MODULE,
1776	.release        = vhost_net_release,
1777	.read_iter      = vhost_net_chr_read_iter,
1778	.write_iter     = vhost_net_chr_write_iter,
1779	.poll           = vhost_net_chr_poll,
1780	.unlocked_ioctl = vhost_net_ioctl,
1781	.compat_ioctl   = compat_ptr_ioctl,
1782	.open           = vhost_net_open,
1783	.llseek		= noop_llseek,
1784};
1785
1786static struct miscdevice vhost_net_misc = {
1787	.minor = VHOST_NET_MINOR,
1788	.name = "vhost-net",
1789	.fops = &vhost_net_fops,
1790};
1791
1792static int vhost_net_init(void)
1793{
1794	if (experimental_zcopytx)
1795		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1796	return misc_register(&vhost_net_misc);
1797}
1798module_init(vhost_net_init);
1799
1800static void vhost_net_exit(void)
1801{
1802	misc_deregister(&vhost_net_misc);
1803}
1804module_exit(vhost_net_exit);
1805
1806MODULE_VERSION("0.0.1");
1807MODULE_LICENSE("GPL v2");
1808MODULE_AUTHOR("Michael S. Tsirkin");
1809MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1810MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1811MODULE_ALIAS("devname:vhost-net");
1812