xref: /linux/drivers/vhost/net.c (revision 63b47f026cc841bd3d3438dd6fccbc394dfead87)
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 
38 static int experimental_zcopytx = 0;
39 module_param(experimental_zcopytx, int, 0444);
40 MODULE_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 
72 enum {
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 			 (1ULL << VIRTIO_F_RING_RESET)
78 };
79 
80 enum {
81 	VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
82 };
83 
84 enum {
85 	VHOST_NET_VQ_RX = 0,
86 	VHOST_NET_VQ_TX = 1,
87 	VHOST_NET_VQ_MAX = 2,
88 };
89 
90 struct vhost_net_ubuf_ref {
91 	/* refcount follows semantics similar to kref:
92 	 *  0: object is released
93 	 *  1: no outstanding ubufs
94 	 * >1: outstanding ubufs
95 	 */
96 	atomic_t refcount;
97 	wait_queue_head_t wait;
98 	struct vhost_virtqueue *vq;
99 };
100 
101 #define VHOST_NET_BATCH 64
102 struct vhost_net_buf {
103 	void **queue;
104 	int tail;
105 	int head;
106 };
107 
108 struct vhost_net_virtqueue {
109 	struct vhost_virtqueue vq;
110 	size_t vhost_hlen;
111 	size_t sock_hlen;
112 	/* vhost zerocopy support fields below: */
113 	/* last used idx for outstanding DMA zerocopy buffers */
114 	int upend_idx;
115 	/* For TX, first used idx for DMA done zerocopy buffers
116 	 * For RX, number of batched heads
117 	 */
118 	int done_idx;
119 	/* Number of XDP frames batched */
120 	int batched_xdp;
121 	/* an array of userspace buffers info */
122 	struct ubuf_info_msgzc *ubuf_info;
123 	/* Reference counting for outstanding ubufs.
124 	 * Protected by vq mutex. Writers must also take device mutex. */
125 	struct vhost_net_ubuf_ref *ubufs;
126 	struct ptr_ring *rx_ring;
127 	struct vhost_net_buf rxq;
128 	/* Batched XDP buffs */
129 	struct xdp_buff *xdp;
130 };
131 
132 struct vhost_net {
133 	struct vhost_dev dev;
134 	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
135 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
136 	/* Number of TX recently submitted.
137 	 * Protected by tx vq lock. */
138 	unsigned tx_packets;
139 	/* Number of times zerocopy TX recently failed.
140 	 * Protected by tx vq lock. */
141 	unsigned tx_zcopy_err;
142 	/* Flush in progress. Protected by tx vq lock. */
143 	bool tx_flush;
144 	/* Private page frag cache */
145 	struct page_frag_cache pf_cache;
146 };
147 
148 static unsigned vhost_net_zcopy_mask __read_mostly;
149 
150 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
151 {
152 	if (rxq->tail != rxq->head)
153 		return rxq->queue[rxq->head];
154 	else
155 		return NULL;
156 }
157 
158 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
159 {
160 	return rxq->tail - rxq->head;
161 }
162 
163 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
164 {
165 	return rxq->tail == rxq->head;
166 }
167 
168 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
169 {
170 	void *ret = vhost_net_buf_get_ptr(rxq);
171 	++rxq->head;
172 	return ret;
173 }
174 
175 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
176 {
177 	struct vhost_net_buf *rxq = &nvq->rxq;
178 
179 	rxq->head = 0;
180 	rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
181 					      VHOST_NET_BATCH);
182 	return rxq->tail;
183 }
184 
185 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
186 {
187 	struct vhost_net_buf *rxq = &nvq->rxq;
188 
189 	if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
190 		ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
191 				   vhost_net_buf_get_size(rxq),
192 				   tun_ptr_free);
193 		rxq->head = rxq->tail = 0;
194 	}
195 }
196 
197 static int vhost_net_buf_peek_len(void *ptr)
198 {
199 	if (tun_is_xdp_frame(ptr)) {
200 		struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
201 
202 		return xdpf->len;
203 	}
204 
205 	return __skb_array_len_with_tag(ptr);
206 }
207 
208 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
209 {
210 	struct vhost_net_buf *rxq = &nvq->rxq;
211 
212 	if (!vhost_net_buf_is_empty(rxq))
213 		goto out;
214 
215 	if (!vhost_net_buf_produce(nvq))
216 		return 0;
217 
218 out:
219 	return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
220 }
221 
222 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
223 {
224 	rxq->head = rxq->tail = 0;
225 }
226 
227 static void vhost_net_enable_zcopy(int vq)
228 {
229 	vhost_net_zcopy_mask |= 0x1 << vq;
230 }
231 
232 static struct vhost_net_ubuf_ref *
233 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
234 {
235 	struct vhost_net_ubuf_ref *ubufs;
236 	/* No zero copy backend? Nothing to count. */
237 	if (!zcopy)
238 		return NULL;
239 	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
240 	if (!ubufs)
241 		return ERR_PTR(-ENOMEM);
242 	atomic_set(&ubufs->refcount, 1);
243 	init_waitqueue_head(&ubufs->wait);
244 	ubufs->vq = vq;
245 	return ubufs;
246 }
247 
248 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
249 {
250 	int r = atomic_sub_return(1, &ubufs->refcount);
251 	if (unlikely(!r))
252 		wake_up(&ubufs->wait);
253 	return r;
254 }
255 
256 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
257 {
258 	vhost_net_ubuf_put(ubufs);
259 	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
260 }
261 
262 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
263 {
264 	vhost_net_ubuf_put_and_wait(ubufs);
265 	kfree(ubufs);
266 }
267 
268 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
269 {
270 	int i;
271 
272 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
273 		kfree(n->vqs[i].ubuf_info);
274 		n->vqs[i].ubuf_info = NULL;
275 	}
276 }
277 
278 static int vhost_net_set_ubuf_info(struct vhost_net *n)
279 {
280 	bool zcopy;
281 	int i;
282 
283 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
284 		zcopy = vhost_net_zcopy_mask & (0x1 << i);
285 		if (!zcopy)
286 			continue;
287 		n->vqs[i].ubuf_info =
288 			kmalloc_array(UIO_MAXIOV,
289 				      sizeof(*n->vqs[i].ubuf_info),
290 				      GFP_KERNEL);
291 		if  (!n->vqs[i].ubuf_info)
292 			goto err;
293 	}
294 	return 0;
295 
296 err:
297 	vhost_net_clear_ubuf_info(n);
298 	return -ENOMEM;
299 }
300 
301 static void vhost_net_vq_reset(struct vhost_net *n)
302 {
303 	int i;
304 
305 	vhost_net_clear_ubuf_info(n);
306 
307 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
308 		n->vqs[i].done_idx = 0;
309 		n->vqs[i].upend_idx = 0;
310 		n->vqs[i].ubufs = NULL;
311 		n->vqs[i].vhost_hlen = 0;
312 		n->vqs[i].sock_hlen = 0;
313 		vhost_net_buf_init(&n->vqs[i].rxq);
314 	}
315 
316 }
317 
318 static void vhost_net_tx_packet(struct vhost_net *net)
319 {
320 	++net->tx_packets;
321 	if (net->tx_packets < 1024)
322 		return;
323 	net->tx_packets = 0;
324 	net->tx_zcopy_err = 0;
325 }
326 
327 static void vhost_net_tx_err(struct vhost_net *net)
328 {
329 	++net->tx_zcopy_err;
330 }
331 
332 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
333 {
334 	/* TX flush waits for outstanding DMAs to be done.
335 	 * Don't start new DMAs.
336 	 */
337 	return !net->tx_flush &&
338 		net->tx_packets / 64 >= net->tx_zcopy_err;
339 }
340 
341 static bool vhost_sock_zcopy(struct socket *sock)
342 {
343 	return unlikely(experimental_zcopytx) &&
344 		sock_flag(sock->sk, SOCK_ZEROCOPY);
345 }
346 
347 static bool vhost_sock_xdp(struct socket *sock)
348 {
349 	return sock_flag(sock->sk, SOCK_XDP);
350 }
351 
352 /* In case of DMA done not in order in lower device driver for some reason.
353  * upend_idx is used to track end of used idx, done_idx is used to track head
354  * of used idx. Once lower device DMA done contiguously, we will signal KVM
355  * guest used idx.
356  */
357 static void vhost_zerocopy_signal_used(struct vhost_net *net,
358 				       struct vhost_virtqueue *vq)
359 {
360 	struct vhost_net_virtqueue *nvq =
361 		container_of(vq, struct vhost_net_virtqueue, vq);
362 	int i, add;
363 	int j = 0;
364 
365 	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
366 		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
367 			vhost_net_tx_err(net);
368 		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
369 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
370 			++j;
371 		} else
372 			break;
373 	}
374 	while (j) {
375 		add = min(UIO_MAXIOV - nvq->done_idx, j);
376 		vhost_add_used_and_signal_n(vq->dev, vq,
377 					    &vq->heads[nvq->done_idx], add);
378 		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
379 		j -= add;
380 	}
381 }
382 
383 static void vhost_zerocopy_complete(struct sk_buff *skb,
384 				    struct ubuf_info *ubuf_base, bool success)
385 {
386 	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
387 	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
388 	struct vhost_virtqueue *vq = ubufs->vq;
389 	int cnt;
390 
391 	rcu_read_lock_bh();
392 
393 	/* set len to mark this desc buffers done DMA */
394 	vq->heads[ubuf->desc].len = success ?
395 		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
396 	cnt = vhost_net_ubuf_put(ubufs);
397 
398 	/*
399 	 * Trigger polling thread if guest stopped submitting new buffers:
400 	 * in this case, the refcount after decrement will eventually reach 1.
401 	 * We also trigger polling periodically after each 16 packets
402 	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
403 	 * less than 10% of times).
404 	 */
405 	if (cnt <= 1 || !(cnt % 16))
406 		vhost_poll_queue(&vq->poll);
407 
408 	rcu_read_unlock_bh();
409 }
410 
411 static const struct ubuf_info_ops vhost_ubuf_ops = {
412 	.complete = vhost_zerocopy_complete,
413 };
414 
415 static inline unsigned long busy_clock(void)
416 {
417 	return local_clock() >> 10;
418 }
419 
420 static bool vhost_can_busy_poll(unsigned long endtime)
421 {
422 	return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
423 		      !signal_pending(current));
424 }
425 
426 static void vhost_net_disable_vq(struct vhost_net *n,
427 				 struct vhost_virtqueue *vq)
428 {
429 	struct vhost_net_virtqueue *nvq =
430 		container_of(vq, struct vhost_net_virtqueue, vq);
431 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
432 	if (!vhost_vq_get_backend(vq))
433 		return;
434 	vhost_poll_stop(poll);
435 }
436 
437 static int vhost_net_enable_vq(struct vhost_net *n,
438 				struct vhost_virtqueue *vq)
439 {
440 	struct vhost_net_virtqueue *nvq =
441 		container_of(vq, struct vhost_net_virtqueue, vq);
442 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
443 	struct socket *sock;
444 
445 	sock = vhost_vq_get_backend(vq);
446 	if (!sock)
447 		return 0;
448 
449 	return vhost_poll_start(poll, sock->file);
450 }
451 
452 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
453 {
454 	struct vhost_virtqueue *vq = &nvq->vq;
455 	struct vhost_dev *dev = vq->dev;
456 
457 	if (!nvq->done_idx)
458 		return;
459 
460 	vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
461 	nvq->done_idx = 0;
462 }
463 
464 static void vhost_tx_batch(struct vhost_net *net,
465 			   struct vhost_net_virtqueue *nvq,
466 			   struct socket *sock,
467 			   struct msghdr *msghdr)
468 {
469 	struct tun_msg_ctl ctl = {
470 		.type = TUN_MSG_PTR,
471 		.num = nvq->batched_xdp,
472 		.ptr = nvq->xdp,
473 	};
474 	int i, err;
475 
476 	if (nvq->batched_xdp == 0)
477 		goto signal_used;
478 
479 	msghdr->msg_control = &ctl;
480 	msghdr->msg_controllen = sizeof(ctl);
481 	err = sock->ops->sendmsg(sock, msghdr, 0);
482 	if (unlikely(err < 0)) {
483 		vq_err(&nvq->vq, "Fail to batch sending packets\n");
484 
485 		/* free pages owned by XDP; since this is an unlikely error path,
486 		 * keep it simple and avoid more complex bulk update for the
487 		 * used pages
488 		 */
489 		for (i = 0; i < nvq->batched_xdp; ++i)
490 			put_page(virt_to_head_page(nvq->xdp[i].data));
491 		nvq->batched_xdp = 0;
492 		nvq->done_idx = 0;
493 		return;
494 	}
495 
496 signal_used:
497 	vhost_net_signal_used(nvq);
498 	nvq->batched_xdp = 0;
499 }
500 
501 static int sock_has_rx_data(struct socket *sock)
502 {
503 	if (unlikely(!sock))
504 		return 0;
505 
506 	if (sock->ops->peek_len)
507 		return sock->ops->peek_len(sock);
508 
509 	return skb_queue_empty(&sock->sk->sk_receive_queue);
510 }
511 
512 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
513 					  struct vhost_virtqueue *vq)
514 {
515 	if (!vhost_vq_avail_empty(&net->dev, vq)) {
516 		vhost_poll_queue(&vq->poll);
517 	} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
518 		vhost_disable_notify(&net->dev, vq);
519 		vhost_poll_queue(&vq->poll);
520 	}
521 }
522 
523 static void vhost_net_busy_poll(struct vhost_net *net,
524 				struct vhost_virtqueue *rvq,
525 				struct vhost_virtqueue *tvq,
526 				bool *busyloop_intr,
527 				bool poll_rx)
528 {
529 	unsigned long busyloop_timeout;
530 	unsigned long endtime;
531 	struct socket *sock;
532 	struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
533 
534 	/* Try to hold the vq mutex of the paired virtqueue. We can't
535 	 * use mutex_lock() here since we could not guarantee a
536 	 * consistenet lock ordering.
537 	 */
538 	if (!mutex_trylock(&vq->mutex))
539 		return;
540 
541 	vhost_disable_notify(&net->dev, vq);
542 	sock = vhost_vq_get_backend(rvq);
543 
544 	busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
545 				     tvq->busyloop_timeout;
546 
547 	preempt_disable();
548 	endtime = busy_clock() + busyloop_timeout;
549 
550 	while (vhost_can_busy_poll(endtime)) {
551 		if (vhost_vq_has_work(vq)) {
552 			*busyloop_intr = true;
553 			break;
554 		}
555 
556 		if ((sock_has_rx_data(sock) &&
557 		     !vhost_vq_avail_empty(&net->dev, rvq)) ||
558 		    !vhost_vq_avail_empty(&net->dev, tvq))
559 			break;
560 
561 		cpu_relax();
562 	}
563 
564 	preempt_enable();
565 
566 	if (poll_rx || sock_has_rx_data(sock))
567 		vhost_net_busy_poll_try_queue(net, vq);
568 	else if (!poll_rx) /* On tx here, sock has no rx data. */
569 		vhost_enable_notify(&net->dev, rvq);
570 
571 	mutex_unlock(&vq->mutex);
572 }
573 
574 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
575 				    struct vhost_net_virtqueue *tnvq,
576 				    unsigned int *out_num, unsigned int *in_num,
577 				    struct msghdr *msghdr, bool *busyloop_intr)
578 {
579 	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
580 	struct vhost_virtqueue *rvq = &rnvq->vq;
581 	struct vhost_virtqueue *tvq = &tnvq->vq;
582 
583 	int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
584 				  out_num, in_num, NULL, NULL);
585 
586 	if (r == tvq->num && tvq->busyloop_timeout) {
587 		/* Flush batched packets first */
588 		if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
589 			vhost_tx_batch(net, tnvq,
590 				       vhost_vq_get_backend(tvq),
591 				       msghdr);
592 
593 		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
594 
595 		r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
596 				      out_num, in_num, NULL, NULL);
597 	}
598 
599 	return r;
600 }
601 
602 static bool vhost_exceeds_maxpend(struct vhost_net *net)
603 {
604 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
605 	struct vhost_virtqueue *vq = &nvq->vq;
606 
607 	return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
608 	       min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
609 }
610 
611 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
612 			    size_t hdr_size, int out)
613 {
614 	/* Skip header. TODO: support TSO. */
615 	size_t len = iov_length(vq->iov, out);
616 
617 	iov_iter_init(iter, ITER_SOURCE, vq->iov, out, len);
618 	iov_iter_advance(iter, hdr_size);
619 
620 	return iov_iter_count(iter);
621 }
622 
623 static int get_tx_bufs(struct vhost_net *net,
624 		       struct vhost_net_virtqueue *nvq,
625 		       struct msghdr *msg,
626 		       unsigned int *out, unsigned int *in,
627 		       size_t *len, bool *busyloop_intr)
628 {
629 	struct vhost_virtqueue *vq = &nvq->vq;
630 	int ret;
631 
632 	ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
633 
634 	if (ret < 0 || ret == vq->num)
635 		return ret;
636 
637 	if (*in) {
638 		vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
639 			*out, *in);
640 		return -EFAULT;
641 	}
642 
643 	/* Sanity check */
644 	*len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
645 	if (*len == 0) {
646 		vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
647 			*len, nvq->vhost_hlen);
648 		return -EFAULT;
649 	}
650 
651 	return ret;
652 }
653 
654 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
655 {
656 	return total_len < VHOST_NET_WEIGHT &&
657 	       !vhost_vq_avail_empty(vq->dev, vq);
658 }
659 
660 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
661 
662 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
663 			       struct iov_iter *from)
664 {
665 	struct vhost_virtqueue *vq = &nvq->vq;
666 	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
667 					     dev);
668 	struct socket *sock = vhost_vq_get_backend(vq);
669 	struct virtio_net_hdr *gso;
670 	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
671 	struct tun_xdp_hdr *hdr;
672 	size_t len = iov_iter_count(from);
673 	int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
674 	int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
675 	int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
676 	int sock_hlen = nvq->sock_hlen;
677 	void *buf;
678 	int copied;
679 	int ret;
680 
681 	if (unlikely(len < nvq->sock_hlen))
682 		return -EFAULT;
683 
684 	if (SKB_DATA_ALIGN(len + pad) +
685 	    SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
686 		return -ENOSPC;
687 
688 	buflen += SKB_DATA_ALIGN(len + pad);
689 	buf = page_frag_alloc_align(&net->pf_cache, buflen, GFP_KERNEL,
690 				    SMP_CACHE_BYTES);
691 	if (unlikely(!buf))
692 		return -ENOMEM;
693 
694 	copied = copy_from_iter(buf + offsetof(struct tun_xdp_hdr, gso),
695 				sock_hlen, from);
696 	if (copied != sock_hlen) {
697 		ret = -EFAULT;
698 		goto err;
699 	}
700 
701 	hdr = buf;
702 	gso = &hdr->gso;
703 
704 	if (!sock_hlen)
705 		memset(buf, 0, pad);
706 
707 	if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
708 	    vhost16_to_cpu(vq, gso->csum_start) +
709 	    vhost16_to_cpu(vq, gso->csum_offset) + 2 >
710 	    vhost16_to_cpu(vq, gso->hdr_len)) {
711 		gso->hdr_len = cpu_to_vhost16(vq,
712 			       vhost16_to_cpu(vq, gso->csum_start) +
713 			       vhost16_to_cpu(vq, gso->csum_offset) + 2);
714 
715 		if (vhost16_to_cpu(vq, gso->hdr_len) > len) {
716 			ret = -EINVAL;
717 			goto err;
718 		}
719 	}
720 
721 	len -= sock_hlen;
722 	copied = copy_from_iter(buf + pad, len, from);
723 	if (copied != len) {
724 		ret = -EFAULT;
725 		goto err;
726 	}
727 
728 	xdp_init_buff(xdp, buflen, NULL);
729 	xdp_prepare_buff(xdp, buf, pad, len, true);
730 	hdr->buflen = buflen;
731 
732 	++nvq->batched_xdp;
733 
734 	return 0;
735 
736 err:
737 	page_frag_free(buf);
738 	return ret;
739 }
740 
741 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
742 {
743 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
744 	struct vhost_virtqueue *vq = &nvq->vq;
745 	unsigned out, in;
746 	int head;
747 	struct msghdr msg = {
748 		.msg_name = NULL,
749 		.msg_namelen = 0,
750 		.msg_control = NULL,
751 		.msg_controllen = 0,
752 		.msg_flags = MSG_DONTWAIT,
753 	};
754 	size_t len, total_len = 0;
755 	int err;
756 	int sent_pkts = 0;
757 	bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
758 
759 	do {
760 		bool busyloop_intr = false;
761 
762 		if (nvq->done_idx == VHOST_NET_BATCH)
763 			vhost_tx_batch(net, nvq, sock, &msg);
764 
765 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
766 				   &busyloop_intr);
767 		/* On error, stop handling until the next kick. */
768 		if (unlikely(head < 0))
769 			break;
770 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
771 		if (head == vq->num) {
772 			if (unlikely(busyloop_intr)) {
773 				vhost_poll_queue(&vq->poll);
774 			} else if (unlikely(vhost_enable_notify(&net->dev,
775 								vq))) {
776 				vhost_disable_notify(&net->dev, vq);
777 				continue;
778 			}
779 			break;
780 		}
781 
782 		total_len += len;
783 
784 		/* For simplicity, TX batching is only enabled if
785 		 * sndbuf is unlimited.
786 		 */
787 		if (sock_can_batch) {
788 			err = vhost_net_build_xdp(nvq, &msg.msg_iter);
789 			if (!err) {
790 				goto done;
791 			} else if (unlikely(err != -ENOSPC)) {
792 				vhost_tx_batch(net, nvq, sock, &msg);
793 				vhost_discard_vq_desc(vq, 1);
794 				vhost_net_enable_vq(net, vq);
795 				break;
796 			}
797 
798 			/* We can't build XDP buff, go for single
799 			 * packet path but let's flush batched
800 			 * packets.
801 			 */
802 			vhost_tx_batch(net, nvq, sock, &msg);
803 			msg.msg_control = NULL;
804 		} else {
805 			if (tx_can_batch(vq, total_len))
806 				msg.msg_flags |= MSG_MORE;
807 			else
808 				msg.msg_flags &= ~MSG_MORE;
809 		}
810 
811 		err = sock->ops->sendmsg(sock, &msg, len);
812 		if (unlikely(err < 0)) {
813 			if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
814 				vhost_discard_vq_desc(vq, 1);
815 				vhost_net_enable_vq(net, vq);
816 				break;
817 			}
818 			pr_debug("Fail to send packet: err %d", err);
819 		} else if (unlikely(err != len))
820 			pr_debug("Truncated TX packet: len %d != %zd\n",
821 				 err, len);
822 done:
823 		vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
824 		vq->heads[nvq->done_idx].len = 0;
825 		++nvq->done_idx;
826 	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
827 
828 	vhost_tx_batch(net, nvq, sock, &msg);
829 }
830 
831 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
832 {
833 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
834 	struct vhost_virtqueue *vq = &nvq->vq;
835 	unsigned out, in;
836 	int head;
837 	struct msghdr msg = {
838 		.msg_name = NULL,
839 		.msg_namelen = 0,
840 		.msg_control = NULL,
841 		.msg_controllen = 0,
842 		.msg_flags = MSG_DONTWAIT,
843 	};
844 	struct tun_msg_ctl ctl;
845 	size_t len, total_len = 0;
846 	int err;
847 	struct vhost_net_ubuf_ref *ubufs;
848 	struct ubuf_info_msgzc *ubuf;
849 	bool zcopy_used;
850 	int sent_pkts = 0;
851 
852 	do {
853 		bool busyloop_intr;
854 
855 		/* Release DMAs done buffers first */
856 		vhost_zerocopy_signal_used(net, vq);
857 
858 		busyloop_intr = false;
859 		head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
860 				   &busyloop_intr);
861 		/* On error, stop handling until the next kick. */
862 		if (unlikely(head < 0))
863 			break;
864 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
865 		if (head == vq->num) {
866 			if (unlikely(busyloop_intr)) {
867 				vhost_poll_queue(&vq->poll);
868 			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
869 				vhost_disable_notify(&net->dev, vq);
870 				continue;
871 			}
872 			break;
873 		}
874 
875 		zcopy_used = len >= VHOST_GOODCOPY_LEN
876 			     && !vhost_exceeds_maxpend(net)
877 			     && vhost_net_tx_select_zcopy(net);
878 
879 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
880 		if (zcopy_used) {
881 			ubuf = nvq->ubuf_info + nvq->upend_idx;
882 			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
883 			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
884 			ubuf->ctx = nvq->ubufs;
885 			ubuf->desc = nvq->upend_idx;
886 			ubuf->ubuf.ops = &vhost_ubuf_ops;
887 			ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
888 			refcount_set(&ubuf->ubuf.refcnt, 1);
889 			msg.msg_control = &ctl;
890 			ctl.type = TUN_MSG_UBUF;
891 			ctl.ptr = &ubuf->ubuf;
892 			msg.msg_controllen = sizeof(ctl);
893 			ubufs = nvq->ubufs;
894 			atomic_inc(&ubufs->refcount);
895 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
896 		} else {
897 			msg.msg_control = NULL;
898 			ubufs = NULL;
899 		}
900 		total_len += len;
901 		if (tx_can_batch(vq, total_len) &&
902 		    likely(!vhost_exceeds_maxpend(net))) {
903 			msg.msg_flags |= MSG_MORE;
904 		} else {
905 			msg.msg_flags &= ~MSG_MORE;
906 		}
907 
908 		err = sock->ops->sendmsg(sock, &msg, len);
909 		if (unlikely(err < 0)) {
910 			bool retry = err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS;
911 
912 			if (zcopy_used) {
913 				if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
914 					vhost_net_ubuf_put(ubufs);
915 				if (retry)
916 					nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
917 						% UIO_MAXIOV;
918 				else
919 					vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
920 			}
921 			if (retry) {
922 				vhost_discard_vq_desc(vq, 1);
923 				vhost_net_enable_vq(net, vq);
924 				break;
925 			}
926 			pr_debug("Fail to send packet: err %d", err);
927 		} else if (unlikely(err != len))
928 			pr_debug("Truncated TX packet: "
929 				 " len %d != %zd\n", err, len);
930 		if (!zcopy_used)
931 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
932 		else
933 			vhost_zerocopy_signal_used(net, vq);
934 		vhost_net_tx_packet(net);
935 	} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
936 }
937 
938 /* Expects to be always run from workqueue - which acts as
939  * read-size critical section for our kind of RCU. */
940 static void handle_tx(struct vhost_net *net)
941 {
942 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
943 	struct vhost_virtqueue *vq = &nvq->vq;
944 	struct socket *sock;
945 
946 	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
947 	sock = vhost_vq_get_backend(vq);
948 	if (!sock)
949 		goto out;
950 
951 	if (!vq_meta_prefetch(vq))
952 		goto out;
953 
954 	vhost_disable_notify(&net->dev, vq);
955 	vhost_net_disable_vq(net, vq);
956 
957 	if (vhost_sock_zcopy(sock))
958 		handle_tx_zerocopy(net, sock);
959 	else
960 		handle_tx_copy(net, sock);
961 
962 out:
963 	mutex_unlock(&vq->mutex);
964 }
965 
966 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
967 {
968 	struct sk_buff *head;
969 	int len = 0;
970 	unsigned long flags;
971 
972 	if (rvq->rx_ring)
973 		return vhost_net_buf_peek(rvq);
974 
975 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
976 	head = skb_peek(&sk->sk_receive_queue);
977 	if (likely(head)) {
978 		len = head->len;
979 		if (skb_vlan_tag_present(head))
980 			len += VLAN_HLEN;
981 	}
982 
983 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
984 	return len;
985 }
986 
987 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
988 				      bool *busyloop_intr)
989 {
990 	struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
991 	struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
992 	struct vhost_virtqueue *rvq = &rnvq->vq;
993 	struct vhost_virtqueue *tvq = &tnvq->vq;
994 	int len = peek_head_len(rnvq, sk);
995 
996 	if (!len && rvq->busyloop_timeout) {
997 		/* Flush batched heads first */
998 		vhost_net_signal_used(rnvq);
999 		/* Both tx vq and rx socket were polled here */
1000 		vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1001 
1002 		len = peek_head_len(rnvq, sk);
1003 	}
1004 
1005 	return len;
1006 }
1007 
1008 /* This is a multi-buffer version of vhost_get_desc, that works if
1009  *	vq has read descriptors only.
1010  * @vq		- the relevant virtqueue
1011  * @datalen	- data length we'll be reading
1012  * @iovcount	- returned count of io vectors we fill
1013  * @log		- vhost log
1014  * @log_num	- log offset
1015  * @quota       - headcount quota, 1 for big buffer
1016  *	returns number of buffer heads allocated, negative on error
1017  */
1018 static int get_rx_bufs(struct vhost_virtqueue *vq,
1019 		       struct vring_used_elem *heads,
1020 		       int datalen,
1021 		       unsigned *iovcount,
1022 		       struct vhost_log *log,
1023 		       unsigned *log_num,
1024 		       unsigned int quota)
1025 {
1026 	unsigned int out, in;
1027 	int seg = 0;
1028 	int headcount = 0;
1029 	unsigned d;
1030 	int r, nlogs = 0;
1031 	/* len is always initialized before use since we are always called with
1032 	 * datalen > 0.
1033 	 */
1034 	u32 len;
1035 
1036 	while (datalen > 0 && headcount < quota) {
1037 		if (unlikely(seg >= UIO_MAXIOV)) {
1038 			r = -ENOBUFS;
1039 			goto err;
1040 		}
1041 		r = vhost_get_vq_desc(vq, vq->iov + seg,
1042 				      ARRAY_SIZE(vq->iov) - seg, &out,
1043 				      &in, log, log_num);
1044 		if (unlikely(r < 0))
1045 			goto err;
1046 
1047 		d = r;
1048 		if (d == vq->num) {
1049 			r = 0;
1050 			goto err;
1051 		}
1052 		if (unlikely(out || in <= 0)) {
1053 			vq_err(vq, "unexpected descriptor format for RX: "
1054 				"out %d, in %d\n", out, in);
1055 			r = -EINVAL;
1056 			goto err;
1057 		}
1058 		if (unlikely(log)) {
1059 			nlogs += *log_num;
1060 			log += *log_num;
1061 		}
1062 		heads[headcount].id = cpu_to_vhost32(vq, d);
1063 		len = iov_length(vq->iov + seg, in);
1064 		heads[headcount].len = cpu_to_vhost32(vq, len);
1065 		datalen -= len;
1066 		++headcount;
1067 		seg += in;
1068 	}
1069 	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1070 	*iovcount = seg;
1071 	if (unlikely(log))
1072 		*log_num = nlogs;
1073 
1074 	/* Detect overrun */
1075 	if (unlikely(datalen > 0)) {
1076 		r = UIO_MAXIOV + 1;
1077 		goto err;
1078 	}
1079 	return headcount;
1080 err:
1081 	vhost_discard_vq_desc(vq, headcount);
1082 	return r;
1083 }
1084 
1085 /* Expects to be always run from workqueue - which acts as
1086  * read-size critical section for our kind of RCU. */
1087 static void handle_rx(struct vhost_net *net)
1088 {
1089 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1090 	struct vhost_virtqueue *vq = &nvq->vq;
1091 	unsigned in, log;
1092 	struct vhost_log *vq_log;
1093 	struct msghdr msg = {
1094 		.msg_name = NULL,
1095 		.msg_namelen = 0,
1096 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
1097 		.msg_controllen = 0,
1098 		.msg_flags = MSG_DONTWAIT,
1099 	};
1100 	struct virtio_net_hdr hdr = {
1101 		.flags = 0,
1102 		.gso_type = VIRTIO_NET_HDR_GSO_NONE
1103 	};
1104 	size_t total_len = 0;
1105 	int err, mergeable;
1106 	s16 headcount;
1107 	size_t vhost_hlen, sock_hlen;
1108 	size_t vhost_len, sock_len;
1109 	bool busyloop_intr = false;
1110 	struct socket *sock;
1111 	struct iov_iter fixup;
1112 	__virtio16 num_buffers;
1113 	int recv_pkts = 0;
1114 
1115 	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1116 	sock = vhost_vq_get_backend(vq);
1117 	if (!sock)
1118 		goto out;
1119 
1120 	if (!vq_meta_prefetch(vq))
1121 		goto out;
1122 
1123 	vhost_disable_notify(&net->dev, vq);
1124 	vhost_net_disable_vq(net, vq);
1125 
1126 	vhost_hlen = nvq->vhost_hlen;
1127 	sock_hlen = nvq->sock_hlen;
1128 
1129 	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1130 		vq->log : NULL;
1131 	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1132 
1133 	do {
1134 		sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1135 						      &busyloop_intr);
1136 		if (!sock_len)
1137 			break;
1138 		sock_len += sock_hlen;
1139 		vhost_len = sock_len + vhost_hlen;
1140 		headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1141 					vhost_len, &in, vq_log, &log,
1142 					likely(mergeable) ? UIO_MAXIOV : 1);
1143 		/* On error, stop handling until the next kick. */
1144 		if (unlikely(headcount < 0))
1145 			goto out;
1146 		/* OK, now we need to know about added descriptors. */
1147 		if (!headcount) {
1148 			if (unlikely(busyloop_intr)) {
1149 				vhost_poll_queue(&vq->poll);
1150 			} else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1151 				/* They have slipped one in as we were
1152 				 * doing that: check again. */
1153 				vhost_disable_notify(&net->dev, vq);
1154 				continue;
1155 			}
1156 			/* Nothing new?  Wait for eventfd to tell us
1157 			 * they refilled. */
1158 			goto out;
1159 		}
1160 		busyloop_intr = false;
1161 		if (nvq->rx_ring)
1162 			msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1163 		/* On overrun, truncate and discard */
1164 		if (unlikely(headcount > UIO_MAXIOV)) {
1165 			iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, 1, 1);
1166 			err = sock->ops->recvmsg(sock, &msg,
1167 						 1, MSG_DONTWAIT | MSG_TRUNC);
1168 			pr_debug("Discarded rx packet: len %zd\n", sock_len);
1169 			continue;
1170 		}
1171 		/* We don't need to be notified again. */
1172 		iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, in, vhost_len);
1173 		fixup = msg.msg_iter;
1174 		if (unlikely((vhost_hlen))) {
1175 			/* We will supply the header ourselves
1176 			 * TODO: support TSO.
1177 			 */
1178 			iov_iter_advance(&msg.msg_iter, vhost_hlen);
1179 		}
1180 		err = sock->ops->recvmsg(sock, &msg,
1181 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1182 		/* Userspace might have consumed the packet meanwhile:
1183 		 * it's not supposed to do this usually, but might be hard
1184 		 * to prevent. Discard data we got (if any) and keep going. */
1185 		if (unlikely(err != sock_len)) {
1186 			pr_debug("Discarded rx packet: "
1187 				 " len %d, expected %zd\n", err, sock_len);
1188 			vhost_discard_vq_desc(vq, headcount);
1189 			continue;
1190 		}
1191 		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1192 		if (unlikely(vhost_hlen)) {
1193 			if (copy_to_iter(&hdr, sizeof(hdr),
1194 					 &fixup) != sizeof(hdr)) {
1195 				vq_err(vq, "Unable to write vnet_hdr "
1196 				       "at addr %p\n", vq->iov->iov_base);
1197 				goto out;
1198 			}
1199 		} else {
1200 			/* Header came from socket; we'll need to patch
1201 			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1202 			 */
1203 			iov_iter_advance(&fixup, sizeof(hdr));
1204 		}
1205 		/* TODO: Should check and handle checksum. */
1206 
1207 		num_buffers = cpu_to_vhost16(vq, headcount);
1208 		if (likely(mergeable) &&
1209 		    copy_to_iter(&num_buffers, sizeof num_buffers,
1210 				 &fixup) != sizeof num_buffers) {
1211 			vq_err(vq, "Failed num_buffers write");
1212 			vhost_discard_vq_desc(vq, headcount);
1213 			goto out;
1214 		}
1215 		nvq->done_idx += headcount;
1216 		if (nvq->done_idx > VHOST_NET_BATCH)
1217 			vhost_net_signal_used(nvq);
1218 		if (unlikely(vq_log))
1219 			vhost_log_write(vq, vq_log, log, vhost_len,
1220 					vq->iov, in);
1221 		total_len += vhost_len;
1222 	} while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1223 
1224 	if (unlikely(busyloop_intr))
1225 		vhost_poll_queue(&vq->poll);
1226 	else if (!sock_len)
1227 		vhost_net_enable_vq(net, vq);
1228 out:
1229 	vhost_net_signal_used(nvq);
1230 	mutex_unlock(&vq->mutex);
1231 }
1232 
1233 static void handle_tx_kick(struct vhost_work *work)
1234 {
1235 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1236 						  poll.work);
1237 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1238 
1239 	handle_tx(net);
1240 }
1241 
1242 static void handle_rx_kick(struct vhost_work *work)
1243 {
1244 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1245 						  poll.work);
1246 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1247 
1248 	handle_rx(net);
1249 }
1250 
1251 static void handle_tx_net(struct vhost_work *work)
1252 {
1253 	struct vhost_net *net = container_of(work, struct vhost_net,
1254 					     poll[VHOST_NET_VQ_TX].work);
1255 	handle_tx(net);
1256 }
1257 
1258 static void handle_rx_net(struct vhost_work *work)
1259 {
1260 	struct vhost_net *net = container_of(work, struct vhost_net,
1261 					     poll[VHOST_NET_VQ_RX].work);
1262 	handle_rx(net);
1263 }
1264 
1265 static int vhost_net_open(struct inode *inode, struct file *f)
1266 {
1267 	struct vhost_net *n;
1268 	struct vhost_dev *dev;
1269 	struct vhost_virtqueue **vqs;
1270 	void **queue;
1271 	struct xdp_buff *xdp;
1272 	int i;
1273 
1274 	n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1275 	if (!n)
1276 		return -ENOMEM;
1277 	vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1278 	if (!vqs) {
1279 		kvfree(n);
1280 		return -ENOMEM;
1281 	}
1282 
1283 	queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1284 			      GFP_KERNEL);
1285 	if (!queue) {
1286 		kfree(vqs);
1287 		kvfree(n);
1288 		return -ENOMEM;
1289 	}
1290 	n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1291 
1292 	xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1293 	if (!xdp) {
1294 		kfree(vqs);
1295 		kvfree(n);
1296 		kfree(queue);
1297 		return -ENOMEM;
1298 	}
1299 	n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1300 
1301 	dev = &n->dev;
1302 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1303 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1304 	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1305 	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1306 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1307 		n->vqs[i].ubufs = NULL;
1308 		n->vqs[i].ubuf_info = NULL;
1309 		n->vqs[i].upend_idx = 0;
1310 		n->vqs[i].done_idx = 0;
1311 		n->vqs[i].batched_xdp = 0;
1312 		n->vqs[i].vhost_hlen = 0;
1313 		n->vqs[i].sock_hlen = 0;
1314 		n->vqs[i].rx_ring = NULL;
1315 		vhost_net_buf_init(&n->vqs[i].rxq);
1316 	}
1317 	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1318 		       UIO_MAXIOV + VHOST_NET_BATCH,
1319 		       VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1320 		       NULL);
1321 
1322 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev,
1323 			vqs[VHOST_NET_VQ_TX]);
1324 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev,
1325 			vqs[VHOST_NET_VQ_RX]);
1326 
1327 	f->private_data = n;
1328 	n->pf_cache.va = NULL;
1329 
1330 	return 0;
1331 }
1332 
1333 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1334 					struct vhost_virtqueue *vq)
1335 {
1336 	struct socket *sock;
1337 	struct vhost_net_virtqueue *nvq =
1338 		container_of(vq, struct vhost_net_virtqueue, vq);
1339 
1340 	mutex_lock(&vq->mutex);
1341 	sock = vhost_vq_get_backend(vq);
1342 	vhost_net_disable_vq(n, vq);
1343 	vhost_vq_set_backend(vq, NULL);
1344 	vhost_net_buf_unproduce(nvq);
1345 	nvq->rx_ring = NULL;
1346 	mutex_unlock(&vq->mutex);
1347 	return sock;
1348 }
1349 
1350 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1351 			   struct socket **rx_sock)
1352 {
1353 	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1354 	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1355 }
1356 
1357 static void vhost_net_flush(struct vhost_net *n)
1358 {
1359 	vhost_dev_flush(&n->dev);
1360 	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1361 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1362 		n->tx_flush = true;
1363 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1364 		/* Wait for all lower device DMAs done. */
1365 		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1366 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1367 		n->tx_flush = false;
1368 		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1369 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1370 	}
1371 }
1372 
1373 static int vhost_net_release(struct inode *inode, struct file *f)
1374 {
1375 	struct vhost_net *n = f->private_data;
1376 	struct socket *tx_sock;
1377 	struct socket *rx_sock;
1378 
1379 	vhost_net_stop(n, &tx_sock, &rx_sock);
1380 	vhost_net_flush(n);
1381 	vhost_dev_stop(&n->dev);
1382 	vhost_dev_cleanup(&n->dev);
1383 	vhost_net_vq_reset(n);
1384 	if (tx_sock)
1385 		sockfd_put(tx_sock);
1386 	if (rx_sock)
1387 		sockfd_put(rx_sock);
1388 	/* Make sure no callbacks are outstanding */
1389 	synchronize_rcu();
1390 	/* We do an extra flush before freeing memory,
1391 	 * since jobs can re-queue themselves. */
1392 	vhost_net_flush(n);
1393 	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1394 	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1395 	kfree(n->dev.vqs);
1396 	page_frag_cache_drain(&n->pf_cache);
1397 	kvfree(n);
1398 	return 0;
1399 }
1400 
1401 static struct socket *get_raw_socket(int fd)
1402 {
1403 	int r;
1404 	struct socket *sock = sockfd_lookup(fd, &r);
1405 
1406 	if (!sock)
1407 		return ERR_PTR(-ENOTSOCK);
1408 
1409 	/* Parameter checking */
1410 	if (sock->sk->sk_type != SOCK_RAW) {
1411 		r = -ESOCKTNOSUPPORT;
1412 		goto err;
1413 	}
1414 
1415 	if (sock->sk->sk_family != AF_PACKET) {
1416 		r = -EPFNOSUPPORT;
1417 		goto err;
1418 	}
1419 	return sock;
1420 err:
1421 	sockfd_put(sock);
1422 	return ERR_PTR(r);
1423 }
1424 
1425 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1426 {
1427 	struct ptr_ring *ring;
1428 	ring = tun_get_tx_ring(file);
1429 	if (!IS_ERR(ring))
1430 		goto out;
1431 	ring = tap_get_ptr_ring(file);
1432 	if (!IS_ERR(ring))
1433 		goto out;
1434 	ring = NULL;
1435 out:
1436 	return ring;
1437 }
1438 
1439 static struct socket *get_tap_socket(int fd)
1440 {
1441 	struct file *file = fget(fd);
1442 	struct socket *sock;
1443 
1444 	if (!file)
1445 		return ERR_PTR(-EBADF);
1446 	sock = tun_get_socket(file);
1447 	if (!IS_ERR(sock))
1448 		return sock;
1449 	sock = tap_get_socket(file);
1450 	if (IS_ERR(sock))
1451 		fput(file);
1452 	return sock;
1453 }
1454 
1455 static struct socket *get_socket(int fd)
1456 {
1457 	struct socket *sock;
1458 
1459 	/* special case to disable backend */
1460 	if (fd == -1)
1461 		return NULL;
1462 	sock = get_raw_socket(fd);
1463 	if (!IS_ERR(sock))
1464 		return sock;
1465 	sock = get_tap_socket(fd);
1466 	if (!IS_ERR(sock))
1467 		return sock;
1468 	return ERR_PTR(-ENOTSOCK);
1469 }
1470 
1471 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1472 {
1473 	struct socket *sock, *oldsock;
1474 	struct vhost_virtqueue *vq;
1475 	struct vhost_net_virtqueue *nvq;
1476 	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1477 	int r;
1478 
1479 	mutex_lock(&n->dev.mutex);
1480 	r = vhost_dev_check_owner(&n->dev);
1481 	if (r)
1482 		goto err;
1483 
1484 	if (index >= VHOST_NET_VQ_MAX) {
1485 		r = -ENOBUFS;
1486 		goto err;
1487 	}
1488 	vq = &n->vqs[index].vq;
1489 	nvq = &n->vqs[index];
1490 	mutex_lock(&vq->mutex);
1491 
1492 	if (fd == -1)
1493 		vhost_clear_msg(&n->dev);
1494 
1495 	/* Verify that ring has been setup correctly. */
1496 	if (!vhost_vq_access_ok(vq)) {
1497 		r = -EFAULT;
1498 		goto err_vq;
1499 	}
1500 	sock = get_socket(fd);
1501 	if (IS_ERR(sock)) {
1502 		r = PTR_ERR(sock);
1503 		goto err_vq;
1504 	}
1505 
1506 	/* start polling new socket */
1507 	oldsock = vhost_vq_get_backend(vq);
1508 	if (sock != oldsock) {
1509 		ubufs = vhost_net_ubuf_alloc(vq,
1510 					     sock && vhost_sock_zcopy(sock));
1511 		if (IS_ERR(ubufs)) {
1512 			r = PTR_ERR(ubufs);
1513 			goto err_ubufs;
1514 		}
1515 
1516 		vhost_net_disable_vq(n, vq);
1517 		vhost_vq_set_backend(vq, sock);
1518 		vhost_net_buf_unproduce(nvq);
1519 		r = vhost_vq_init_access(vq);
1520 		if (r)
1521 			goto err_used;
1522 		r = vhost_net_enable_vq(n, vq);
1523 		if (r)
1524 			goto err_used;
1525 		if (index == VHOST_NET_VQ_RX) {
1526 			if (sock)
1527 				nvq->rx_ring = get_tap_ptr_ring(sock->file);
1528 			else
1529 				nvq->rx_ring = NULL;
1530 		}
1531 
1532 		oldubufs = nvq->ubufs;
1533 		nvq->ubufs = ubufs;
1534 
1535 		n->tx_packets = 0;
1536 		n->tx_zcopy_err = 0;
1537 		n->tx_flush = false;
1538 	}
1539 
1540 	mutex_unlock(&vq->mutex);
1541 
1542 	if (oldubufs) {
1543 		vhost_net_ubuf_put_wait_and_free(oldubufs);
1544 		mutex_lock(&vq->mutex);
1545 		vhost_zerocopy_signal_used(n, vq);
1546 		mutex_unlock(&vq->mutex);
1547 	}
1548 
1549 	if (oldsock) {
1550 		vhost_dev_flush(&n->dev);
1551 		sockfd_put(oldsock);
1552 	}
1553 
1554 	mutex_unlock(&n->dev.mutex);
1555 	return 0;
1556 
1557 err_used:
1558 	vhost_vq_set_backend(vq, oldsock);
1559 	vhost_net_enable_vq(n, vq);
1560 	if (ubufs)
1561 		vhost_net_ubuf_put_wait_and_free(ubufs);
1562 err_ubufs:
1563 	if (sock)
1564 		sockfd_put(sock);
1565 err_vq:
1566 	mutex_unlock(&vq->mutex);
1567 err:
1568 	mutex_unlock(&n->dev.mutex);
1569 	return r;
1570 }
1571 
1572 static long vhost_net_reset_owner(struct vhost_net *n)
1573 {
1574 	struct socket *tx_sock = NULL;
1575 	struct socket *rx_sock = NULL;
1576 	long err;
1577 	struct vhost_iotlb *umem;
1578 
1579 	mutex_lock(&n->dev.mutex);
1580 	err = vhost_dev_check_owner(&n->dev);
1581 	if (err)
1582 		goto done;
1583 	umem = vhost_dev_reset_owner_prepare();
1584 	if (!umem) {
1585 		err = -ENOMEM;
1586 		goto done;
1587 	}
1588 	vhost_net_stop(n, &tx_sock, &rx_sock);
1589 	vhost_net_flush(n);
1590 	vhost_dev_stop(&n->dev);
1591 	vhost_dev_reset_owner(&n->dev, umem);
1592 	vhost_net_vq_reset(n);
1593 done:
1594 	mutex_unlock(&n->dev.mutex);
1595 	if (tx_sock)
1596 		sockfd_put(tx_sock);
1597 	if (rx_sock)
1598 		sockfd_put(rx_sock);
1599 	return err;
1600 }
1601 
1602 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1603 {
1604 	size_t vhost_hlen, sock_hlen, hdr_len;
1605 	int i;
1606 
1607 	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1608 			       (1ULL << VIRTIO_F_VERSION_1))) ?
1609 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1610 			sizeof(struct virtio_net_hdr);
1611 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1612 		/* vhost provides vnet_hdr */
1613 		vhost_hlen = hdr_len;
1614 		sock_hlen = 0;
1615 	} else {
1616 		/* socket provides vnet_hdr */
1617 		vhost_hlen = 0;
1618 		sock_hlen = hdr_len;
1619 	}
1620 	mutex_lock(&n->dev.mutex);
1621 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1622 	    !vhost_log_access_ok(&n->dev))
1623 		goto out_unlock;
1624 
1625 	if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1626 		if (vhost_init_device_iotlb(&n->dev))
1627 			goto out_unlock;
1628 	}
1629 
1630 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1631 		mutex_lock(&n->vqs[i].vq.mutex);
1632 		n->vqs[i].vq.acked_features = features;
1633 		n->vqs[i].vhost_hlen = vhost_hlen;
1634 		n->vqs[i].sock_hlen = sock_hlen;
1635 		mutex_unlock(&n->vqs[i].vq.mutex);
1636 	}
1637 	mutex_unlock(&n->dev.mutex);
1638 	return 0;
1639 
1640 out_unlock:
1641 	mutex_unlock(&n->dev.mutex);
1642 	return -EFAULT;
1643 }
1644 
1645 static long vhost_net_set_owner(struct vhost_net *n)
1646 {
1647 	int r;
1648 
1649 	mutex_lock(&n->dev.mutex);
1650 	if (vhost_dev_has_owner(&n->dev)) {
1651 		r = -EBUSY;
1652 		goto out;
1653 	}
1654 	r = vhost_net_set_ubuf_info(n);
1655 	if (r)
1656 		goto out;
1657 	r = vhost_dev_set_owner(&n->dev);
1658 	if (r)
1659 		vhost_net_clear_ubuf_info(n);
1660 	vhost_net_flush(n);
1661 out:
1662 	mutex_unlock(&n->dev.mutex);
1663 	return r;
1664 }
1665 
1666 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1667 			    unsigned long arg)
1668 {
1669 	struct vhost_net *n = f->private_data;
1670 	void __user *argp = (void __user *)arg;
1671 	u64 __user *featurep = argp;
1672 	struct vhost_vring_file backend;
1673 	u64 features;
1674 	int r;
1675 
1676 	switch (ioctl) {
1677 	case VHOST_NET_SET_BACKEND:
1678 		if (copy_from_user(&backend, argp, sizeof backend))
1679 			return -EFAULT;
1680 		return vhost_net_set_backend(n, backend.index, backend.fd);
1681 	case VHOST_GET_FEATURES:
1682 		features = VHOST_NET_FEATURES;
1683 		if (copy_to_user(featurep, &features, sizeof features))
1684 			return -EFAULT;
1685 		return 0;
1686 	case VHOST_SET_FEATURES:
1687 		if (copy_from_user(&features, featurep, sizeof features))
1688 			return -EFAULT;
1689 		if (features & ~VHOST_NET_FEATURES)
1690 			return -EOPNOTSUPP;
1691 		return vhost_net_set_features(n, features);
1692 	case VHOST_GET_BACKEND_FEATURES:
1693 		features = VHOST_NET_BACKEND_FEATURES;
1694 		if (copy_to_user(featurep, &features, sizeof(features)))
1695 			return -EFAULT;
1696 		return 0;
1697 	case VHOST_SET_BACKEND_FEATURES:
1698 		if (copy_from_user(&features, featurep, sizeof(features)))
1699 			return -EFAULT;
1700 		if (features & ~VHOST_NET_BACKEND_FEATURES)
1701 			return -EOPNOTSUPP;
1702 		vhost_set_backend_features(&n->dev, features);
1703 		return 0;
1704 	case VHOST_RESET_OWNER:
1705 		return vhost_net_reset_owner(n);
1706 	case VHOST_SET_OWNER:
1707 		return vhost_net_set_owner(n);
1708 	default:
1709 		mutex_lock(&n->dev.mutex);
1710 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1711 		if (r == -ENOIOCTLCMD)
1712 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1713 		else
1714 			vhost_net_flush(n);
1715 		mutex_unlock(&n->dev.mutex);
1716 		return r;
1717 	}
1718 }
1719 
1720 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1721 {
1722 	struct file *file = iocb->ki_filp;
1723 	struct vhost_net *n = file->private_data;
1724 	struct vhost_dev *dev = &n->dev;
1725 	int noblock = file->f_flags & O_NONBLOCK;
1726 
1727 	return vhost_chr_read_iter(dev, to, noblock);
1728 }
1729 
1730 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1731 					struct iov_iter *from)
1732 {
1733 	struct file *file = iocb->ki_filp;
1734 	struct vhost_net *n = file->private_data;
1735 	struct vhost_dev *dev = &n->dev;
1736 
1737 	return vhost_chr_write_iter(dev, from);
1738 }
1739 
1740 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1741 {
1742 	struct vhost_net *n = file->private_data;
1743 	struct vhost_dev *dev = &n->dev;
1744 
1745 	return vhost_chr_poll(file, dev, wait);
1746 }
1747 
1748 static const struct file_operations vhost_net_fops = {
1749 	.owner          = THIS_MODULE,
1750 	.release        = vhost_net_release,
1751 	.read_iter      = vhost_net_chr_read_iter,
1752 	.write_iter     = vhost_net_chr_write_iter,
1753 	.poll           = vhost_net_chr_poll,
1754 	.unlocked_ioctl = vhost_net_ioctl,
1755 	.compat_ioctl   = compat_ptr_ioctl,
1756 	.open           = vhost_net_open,
1757 	.llseek		= noop_llseek,
1758 };
1759 
1760 static struct miscdevice vhost_net_misc = {
1761 	.minor = VHOST_NET_MINOR,
1762 	.name = "vhost-net",
1763 	.fops = &vhost_net_fops,
1764 };
1765 
1766 static int __init vhost_net_init(void)
1767 {
1768 	if (experimental_zcopytx)
1769 		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1770 	return misc_register(&vhost_net_misc);
1771 }
1772 module_init(vhost_net_init);
1773 
1774 static void __exit vhost_net_exit(void)
1775 {
1776 	misc_deregister(&vhost_net_misc);
1777 }
1778 module_exit(vhost_net_exit);
1779 
1780 MODULE_VERSION("0.0.1");
1781 MODULE_LICENSE("GPL v2");
1782 MODULE_AUTHOR("Michael S. Tsirkin");
1783 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1784 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1785 MODULE_ALIAS("devname:vhost-net");
1786