xref: /openbmc/linux/drivers/vhost/net.c (revision 4800cd83)
1 /* Copyright (C) 2009 Red Hat, Inc.
2  * Author: Michael S. Tsirkin <mst@redhat.com>
3  *
4  * This work is licensed under the terms of the GNU GPL, version 2.
5  *
6  * virtio-net server in host kernel.
7  */
8 
9 #include <linux/compat.h>
10 #include <linux/eventfd.h>
11 #include <linux/vhost.h>
12 #include <linux/virtio_net.h>
13 #include <linux/miscdevice.h>
14 #include <linux/module.h>
15 #include <linux/mutex.h>
16 #include <linux/workqueue.h>
17 #include <linux/rcupdate.h>
18 #include <linux/file.h>
19 #include <linux/slab.h>
20 
21 #include <linux/net.h>
22 #include <linux/if_packet.h>
23 #include <linux/if_arp.h>
24 #include <linux/if_tun.h>
25 #include <linux/if_macvlan.h>
26 
27 #include <net/sock.h>
28 
29 #include "vhost.h"
30 
31 /* Max number of bytes transferred before requeueing the job.
32  * Using this limit prevents one virtqueue from starving others. */
33 #define VHOST_NET_WEIGHT 0x80000
34 
35 enum {
36 	VHOST_NET_VQ_RX = 0,
37 	VHOST_NET_VQ_TX = 1,
38 	VHOST_NET_VQ_MAX = 2,
39 };
40 
41 enum vhost_net_poll_state {
42 	VHOST_NET_POLL_DISABLED = 0,
43 	VHOST_NET_POLL_STARTED = 1,
44 	VHOST_NET_POLL_STOPPED = 2,
45 };
46 
47 struct vhost_net {
48 	struct vhost_dev dev;
49 	struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
50 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
51 	/* Tells us whether we are polling a socket for TX.
52 	 * We only do this when socket buffer fills up.
53 	 * Protected by tx vq lock. */
54 	enum vhost_net_poll_state tx_poll_state;
55 };
56 
57 /* Pop first len bytes from iovec. Return number of segments used. */
58 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
59 			  size_t len, int iov_count)
60 {
61 	int seg = 0;
62 	size_t size;
63 	while (len && seg < iov_count) {
64 		size = min(from->iov_len, len);
65 		to->iov_base = from->iov_base;
66 		to->iov_len = size;
67 		from->iov_len -= size;
68 		from->iov_base += size;
69 		len -= size;
70 		++from;
71 		++to;
72 		++seg;
73 	}
74 	return seg;
75 }
76 /* Copy iovec entries for len bytes from iovec. */
77 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
78 			   size_t len, int iovcount)
79 {
80 	int seg = 0;
81 	size_t size;
82 	while (len && seg < iovcount) {
83 		size = min(from->iov_len, len);
84 		to->iov_base = from->iov_base;
85 		to->iov_len = size;
86 		len -= size;
87 		++from;
88 		++to;
89 		++seg;
90 	}
91 }
92 
93 /* Caller must have TX VQ lock */
94 static void tx_poll_stop(struct vhost_net *net)
95 {
96 	if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
97 		return;
98 	vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
99 	net->tx_poll_state = VHOST_NET_POLL_STOPPED;
100 }
101 
102 /* Caller must have TX VQ lock */
103 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
104 {
105 	if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
106 		return;
107 	vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
108 	net->tx_poll_state = VHOST_NET_POLL_STARTED;
109 }
110 
111 /* Expects to be always run from workqueue - which acts as
112  * read-size critical section for our kind of RCU. */
113 static void handle_tx(struct vhost_net *net)
114 {
115 	struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
116 	unsigned out, in, s;
117 	int head;
118 	struct msghdr msg = {
119 		.msg_name = NULL,
120 		.msg_namelen = 0,
121 		.msg_control = NULL,
122 		.msg_controllen = 0,
123 		.msg_iov = vq->iov,
124 		.msg_flags = MSG_DONTWAIT,
125 	};
126 	size_t len, total_len = 0;
127 	int err, wmem;
128 	size_t hdr_size;
129 	struct socket *sock;
130 
131 	/* TODO: check that we are running from vhost_worker? */
132 	sock = rcu_dereference_check(vq->private_data, 1);
133 	if (!sock)
134 		return;
135 
136 	wmem = atomic_read(&sock->sk->sk_wmem_alloc);
137 	if (wmem >= sock->sk->sk_sndbuf) {
138 		mutex_lock(&vq->mutex);
139 		tx_poll_start(net, sock);
140 		mutex_unlock(&vq->mutex);
141 		return;
142 	}
143 
144 	mutex_lock(&vq->mutex);
145 	vhost_disable_notify(vq);
146 
147 	if (wmem < sock->sk->sk_sndbuf / 2)
148 		tx_poll_stop(net);
149 	hdr_size = vq->vhost_hlen;
150 
151 	for (;;) {
152 		head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
153 					 ARRAY_SIZE(vq->iov),
154 					 &out, &in,
155 					 NULL, NULL);
156 		/* On error, stop handling until the next kick. */
157 		if (unlikely(head < 0))
158 			break;
159 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
160 		if (head == vq->num) {
161 			wmem = atomic_read(&sock->sk->sk_wmem_alloc);
162 			if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
163 				tx_poll_start(net, sock);
164 				set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
165 				break;
166 			}
167 			if (unlikely(vhost_enable_notify(vq))) {
168 				vhost_disable_notify(vq);
169 				continue;
170 			}
171 			break;
172 		}
173 		if (in) {
174 			vq_err(vq, "Unexpected descriptor format for TX: "
175 			       "out %d, int %d\n", out, in);
176 			break;
177 		}
178 		/* Skip header. TODO: support TSO. */
179 		s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
180 		msg.msg_iovlen = out;
181 		len = iov_length(vq->iov, out);
182 		/* Sanity check */
183 		if (!len) {
184 			vq_err(vq, "Unexpected header len for TX: "
185 			       "%zd expected %zd\n",
186 			       iov_length(vq->hdr, s), hdr_size);
187 			break;
188 		}
189 		/* TODO: Check specific error and bomb out unless ENOBUFS? */
190 		err = sock->ops->sendmsg(NULL, sock, &msg, len);
191 		if (unlikely(err < 0)) {
192 			vhost_discard_vq_desc(vq, 1);
193 			tx_poll_start(net, sock);
194 			break;
195 		}
196 		if (err != len)
197 			pr_debug("Truncated TX packet: "
198 				 " len %d != %zd\n", err, len);
199 		vhost_add_used_and_signal(&net->dev, vq, head, 0);
200 		total_len += len;
201 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
202 			vhost_poll_queue(&vq->poll);
203 			break;
204 		}
205 	}
206 
207 	mutex_unlock(&vq->mutex);
208 }
209 
210 static int peek_head_len(struct sock *sk)
211 {
212 	struct sk_buff *head;
213 	int len = 0;
214 
215 	lock_sock(sk);
216 	head = skb_peek(&sk->sk_receive_queue);
217 	if (head)
218 		len = head->len;
219 	release_sock(sk);
220 	return len;
221 }
222 
223 /* This is a multi-buffer version of vhost_get_desc, that works if
224  *	vq has read descriptors only.
225  * @vq		- the relevant virtqueue
226  * @datalen	- data length we'll be reading
227  * @iovcount	- returned count of io vectors we fill
228  * @log		- vhost log
229  * @log_num	- log offset
230  *	returns number of buffer heads allocated, negative on error
231  */
232 static int get_rx_bufs(struct vhost_virtqueue *vq,
233 		       struct vring_used_elem *heads,
234 		       int datalen,
235 		       unsigned *iovcount,
236 		       struct vhost_log *log,
237 		       unsigned *log_num)
238 {
239 	unsigned int out, in;
240 	int seg = 0;
241 	int headcount = 0;
242 	unsigned d;
243 	int r, nlogs = 0;
244 
245 	while (datalen > 0) {
246 		if (unlikely(seg >= UIO_MAXIOV)) {
247 			r = -ENOBUFS;
248 			goto err;
249 		}
250 		d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
251 				      ARRAY_SIZE(vq->iov) - seg, &out,
252 				      &in, log, log_num);
253 		if (d == vq->num) {
254 			r = 0;
255 			goto err;
256 		}
257 		if (unlikely(out || in <= 0)) {
258 			vq_err(vq, "unexpected descriptor format for RX: "
259 				"out %d, in %d\n", out, in);
260 			r = -EINVAL;
261 			goto err;
262 		}
263 		if (unlikely(log)) {
264 			nlogs += *log_num;
265 			log += *log_num;
266 		}
267 		heads[headcount].id = d;
268 		heads[headcount].len = iov_length(vq->iov + seg, in);
269 		datalen -= heads[headcount].len;
270 		++headcount;
271 		seg += in;
272 	}
273 	heads[headcount - 1].len += datalen;
274 	*iovcount = seg;
275 	if (unlikely(log))
276 		*log_num = nlogs;
277 	return headcount;
278 err:
279 	vhost_discard_vq_desc(vq, headcount);
280 	return r;
281 }
282 
283 /* Expects to be always run from workqueue - which acts as
284  * read-size critical section for our kind of RCU. */
285 static void handle_rx_big(struct vhost_net *net)
286 {
287 	struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
288 	unsigned out, in, log, s;
289 	int head;
290 	struct vhost_log *vq_log;
291 	struct msghdr msg = {
292 		.msg_name = NULL,
293 		.msg_namelen = 0,
294 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
295 		.msg_controllen = 0,
296 		.msg_iov = vq->iov,
297 		.msg_flags = MSG_DONTWAIT,
298 	};
299 
300 	struct virtio_net_hdr hdr = {
301 		.flags = 0,
302 		.gso_type = VIRTIO_NET_HDR_GSO_NONE
303 	};
304 
305 	size_t len, total_len = 0;
306 	int err;
307 	size_t hdr_size;
308 	/* TODO: check that we are running from vhost_worker? */
309 	struct socket *sock = rcu_dereference_check(vq->private_data, 1);
310 	if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
311 		return;
312 
313 	mutex_lock(&vq->mutex);
314 	vhost_disable_notify(vq);
315 	hdr_size = vq->vhost_hlen;
316 
317 	vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
318 		vq->log : NULL;
319 
320 	for (;;) {
321 		head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
322 					 ARRAY_SIZE(vq->iov),
323 					 &out, &in,
324 					 vq_log, &log);
325 		/* On error, stop handling until the next kick. */
326 		if (unlikely(head < 0))
327 			break;
328 		/* OK, now we need to know about added descriptors. */
329 		if (head == vq->num) {
330 			if (unlikely(vhost_enable_notify(vq))) {
331 				/* They have slipped one in as we were
332 				 * doing that: check again. */
333 				vhost_disable_notify(vq);
334 				continue;
335 			}
336 			/* Nothing new?  Wait for eventfd to tell us
337 			 * they refilled. */
338 			break;
339 		}
340 		/* We don't need to be notified again. */
341 		if (out) {
342 			vq_err(vq, "Unexpected descriptor format for RX: "
343 			       "out %d, int %d\n",
344 			       out, in);
345 			break;
346 		}
347 		/* Skip header. TODO: support TSO/mergeable rx buffers. */
348 		s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
349 		msg.msg_iovlen = in;
350 		len = iov_length(vq->iov, in);
351 		/* Sanity check */
352 		if (!len) {
353 			vq_err(vq, "Unexpected header len for RX: "
354 			       "%zd expected %zd\n",
355 			       iov_length(vq->hdr, s), hdr_size);
356 			break;
357 		}
358 		err = sock->ops->recvmsg(NULL, sock, &msg,
359 					 len, MSG_DONTWAIT | MSG_TRUNC);
360 		/* TODO: Check specific error and bomb out unless EAGAIN? */
361 		if (err < 0) {
362 			vhost_discard_vq_desc(vq, 1);
363 			break;
364 		}
365 		/* TODO: Should check and handle checksum. */
366 		if (err > len) {
367 			pr_debug("Discarded truncated rx packet: "
368 				 " len %d > %zd\n", err, len);
369 			vhost_discard_vq_desc(vq, 1);
370 			continue;
371 		}
372 		len = err;
373 		err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
374 		if (err) {
375 			vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
376 			       vq->iov->iov_base, err);
377 			break;
378 		}
379 		len += hdr_size;
380 		vhost_add_used_and_signal(&net->dev, vq, head, len);
381 		if (unlikely(vq_log))
382 			vhost_log_write(vq, vq_log, log, len);
383 		total_len += len;
384 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
385 			vhost_poll_queue(&vq->poll);
386 			break;
387 		}
388 	}
389 
390 	mutex_unlock(&vq->mutex);
391 }
392 
393 /* Expects to be always run from workqueue - which acts as
394  * read-size critical section for our kind of RCU. */
395 static void handle_rx_mergeable(struct vhost_net *net)
396 {
397 	struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
398 	unsigned uninitialized_var(in), log;
399 	struct vhost_log *vq_log;
400 	struct msghdr msg = {
401 		.msg_name = NULL,
402 		.msg_namelen = 0,
403 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
404 		.msg_controllen = 0,
405 		.msg_iov = vq->iov,
406 		.msg_flags = MSG_DONTWAIT,
407 	};
408 
409 	struct virtio_net_hdr_mrg_rxbuf hdr = {
410 		.hdr.flags = 0,
411 		.hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
412 	};
413 
414 	size_t total_len = 0;
415 	int err, headcount;
416 	size_t vhost_hlen, sock_hlen;
417 	size_t vhost_len, sock_len;
418 	/* TODO: check that we are running from vhost_worker? */
419 	struct socket *sock = rcu_dereference_check(vq->private_data, 1);
420 	if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
421 		return;
422 
423 	mutex_lock(&vq->mutex);
424 	vhost_disable_notify(vq);
425 	vhost_hlen = vq->vhost_hlen;
426 	sock_hlen = vq->sock_hlen;
427 
428 	vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
429 		vq->log : NULL;
430 
431 	while ((sock_len = peek_head_len(sock->sk))) {
432 		sock_len += sock_hlen;
433 		vhost_len = sock_len + vhost_hlen;
434 		headcount = get_rx_bufs(vq, vq->heads, vhost_len,
435 					&in, vq_log, &log);
436 		/* On error, stop handling until the next kick. */
437 		if (unlikely(headcount < 0))
438 			break;
439 		/* OK, now we need to know about added descriptors. */
440 		if (!headcount) {
441 			if (unlikely(vhost_enable_notify(vq))) {
442 				/* They have slipped one in as we were
443 				 * doing that: check again. */
444 				vhost_disable_notify(vq);
445 				continue;
446 			}
447 			/* Nothing new?  Wait for eventfd to tell us
448 			 * they refilled. */
449 			break;
450 		}
451 		/* We don't need to be notified again. */
452 		if (unlikely((vhost_hlen)))
453 			/* Skip header. TODO: support TSO. */
454 			move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
455 		else
456 			/* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
457 			 * needed because recvmsg can modify msg_iov. */
458 			copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
459 		msg.msg_iovlen = in;
460 		err = sock->ops->recvmsg(NULL, sock, &msg,
461 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
462 		/* Userspace might have consumed the packet meanwhile:
463 		 * it's not supposed to do this usually, but might be hard
464 		 * to prevent. Discard data we got (if any) and keep going. */
465 		if (unlikely(err != sock_len)) {
466 			pr_debug("Discarded rx packet: "
467 				 " len %d, expected %zd\n", err, sock_len);
468 			vhost_discard_vq_desc(vq, headcount);
469 			continue;
470 		}
471 		if (unlikely(vhost_hlen) &&
472 		    memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
473 				      vhost_hlen)) {
474 			vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
475 			       vq->iov->iov_base);
476 			break;
477 		}
478 		/* TODO: Should check and handle checksum. */
479 		if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF) &&
480 		    memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
481 				      offsetof(typeof(hdr), num_buffers),
482 				      sizeof hdr.num_buffers)) {
483 			vq_err(vq, "Failed num_buffers write");
484 			vhost_discard_vq_desc(vq, headcount);
485 			break;
486 		}
487 		vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
488 					    headcount);
489 		if (unlikely(vq_log))
490 			vhost_log_write(vq, vq_log, log, vhost_len);
491 		total_len += vhost_len;
492 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
493 			vhost_poll_queue(&vq->poll);
494 			break;
495 		}
496 	}
497 
498 	mutex_unlock(&vq->mutex);
499 }
500 
501 static void handle_rx(struct vhost_net *net)
502 {
503 	if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF))
504 		handle_rx_mergeable(net);
505 	else
506 		handle_rx_big(net);
507 }
508 
509 static void handle_tx_kick(struct vhost_work *work)
510 {
511 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
512 						  poll.work);
513 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
514 
515 	handle_tx(net);
516 }
517 
518 static void handle_rx_kick(struct vhost_work *work)
519 {
520 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
521 						  poll.work);
522 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
523 
524 	handle_rx(net);
525 }
526 
527 static void handle_tx_net(struct vhost_work *work)
528 {
529 	struct vhost_net *net = container_of(work, struct vhost_net,
530 					     poll[VHOST_NET_VQ_TX].work);
531 	handle_tx(net);
532 }
533 
534 static void handle_rx_net(struct vhost_work *work)
535 {
536 	struct vhost_net *net = container_of(work, struct vhost_net,
537 					     poll[VHOST_NET_VQ_RX].work);
538 	handle_rx(net);
539 }
540 
541 static int vhost_net_open(struct inode *inode, struct file *f)
542 {
543 	struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
544 	struct vhost_dev *dev;
545 	int r;
546 
547 	if (!n)
548 		return -ENOMEM;
549 
550 	dev = &n->dev;
551 	n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
552 	n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
553 	r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
554 	if (r < 0) {
555 		kfree(n);
556 		return r;
557 	}
558 
559 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
560 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
561 	n->tx_poll_state = VHOST_NET_POLL_DISABLED;
562 
563 	f->private_data = n;
564 
565 	return 0;
566 }
567 
568 static void vhost_net_disable_vq(struct vhost_net *n,
569 				 struct vhost_virtqueue *vq)
570 {
571 	if (!vq->private_data)
572 		return;
573 	if (vq == n->vqs + VHOST_NET_VQ_TX) {
574 		tx_poll_stop(n);
575 		n->tx_poll_state = VHOST_NET_POLL_DISABLED;
576 	} else
577 		vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
578 }
579 
580 static void vhost_net_enable_vq(struct vhost_net *n,
581 				struct vhost_virtqueue *vq)
582 {
583 	struct socket *sock;
584 
585 	sock = rcu_dereference_protected(vq->private_data,
586 					 lockdep_is_held(&vq->mutex));
587 	if (!sock)
588 		return;
589 	if (vq == n->vqs + VHOST_NET_VQ_TX) {
590 		n->tx_poll_state = VHOST_NET_POLL_STOPPED;
591 		tx_poll_start(n, sock);
592 	} else
593 		vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
594 }
595 
596 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
597 					struct vhost_virtqueue *vq)
598 {
599 	struct socket *sock;
600 
601 	mutex_lock(&vq->mutex);
602 	sock = rcu_dereference_protected(vq->private_data,
603 					 lockdep_is_held(&vq->mutex));
604 	vhost_net_disable_vq(n, vq);
605 	rcu_assign_pointer(vq->private_data, NULL);
606 	mutex_unlock(&vq->mutex);
607 	return sock;
608 }
609 
610 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
611 			   struct socket **rx_sock)
612 {
613 	*tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
614 	*rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
615 }
616 
617 static void vhost_net_flush_vq(struct vhost_net *n, int index)
618 {
619 	vhost_poll_flush(n->poll + index);
620 	vhost_poll_flush(&n->dev.vqs[index].poll);
621 }
622 
623 static void vhost_net_flush(struct vhost_net *n)
624 {
625 	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
626 	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
627 }
628 
629 static int vhost_net_release(struct inode *inode, struct file *f)
630 {
631 	struct vhost_net *n = f->private_data;
632 	struct socket *tx_sock;
633 	struct socket *rx_sock;
634 
635 	vhost_net_stop(n, &tx_sock, &rx_sock);
636 	vhost_net_flush(n);
637 	vhost_dev_cleanup(&n->dev);
638 	if (tx_sock)
639 		fput(tx_sock->file);
640 	if (rx_sock)
641 		fput(rx_sock->file);
642 	/* We do an extra flush before freeing memory,
643 	 * since jobs can re-queue themselves. */
644 	vhost_net_flush(n);
645 	kfree(n);
646 	return 0;
647 }
648 
649 static struct socket *get_raw_socket(int fd)
650 {
651 	struct {
652 		struct sockaddr_ll sa;
653 		char  buf[MAX_ADDR_LEN];
654 	} uaddr;
655 	int uaddr_len = sizeof uaddr, r;
656 	struct socket *sock = sockfd_lookup(fd, &r);
657 	if (!sock)
658 		return ERR_PTR(-ENOTSOCK);
659 
660 	/* Parameter checking */
661 	if (sock->sk->sk_type != SOCK_RAW) {
662 		r = -ESOCKTNOSUPPORT;
663 		goto err;
664 	}
665 
666 	r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
667 			       &uaddr_len, 0);
668 	if (r)
669 		goto err;
670 
671 	if (uaddr.sa.sll_family != AF_PACKET) {
672 		r = -EPFNOSUPPORT;
673 		goto err;
674 	}
675 	return sock;
676 err:
677 	fput(sock->file);
678 	return ERR_PTR(r);
679 }
680 
681 static struct socket *get_tap_socket(int fd)
682 {
683 	struct file *file = fget(fd);
684 	struct socket *sock;
685 	if (!file)
686 		return ERR_PTR(-EBADF);
687 	sock = tun_get_socket(file);
688 	if (!IS_ERR(sock))
689 		return sock;
690 	sock = macvtap_get_socket(file);
691 	if (IS_ERR(sock))
692 		fput(file);
693 	return sock;
694 }
695 
696 static struct socket *get_socket(int fd)
697 {
698 	struct socket *sock;
699 	/* special case to disable backend */
700 	if (fd == -1)
701 		return NULL;
702 	sock = get_raw_socket(fd);
703 	if (!IS_ERR(sock))
704 		return sock;
705 	sock = get_tap_socket(fd);
706 	if (!IS_ERR(sock))
707 		return sock;
708 	return ERR_PTR(-ENOTSOCK);
709 }
710 
711 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
712 {
713 	struct socket *sock, *oldsock;
714 	struct vhost_virtqueue *vq;
715 	int r;
716 
717 	mutex_lock(&n->dev.mutex);
718 	r = vhost_dev_check_owner(&n->dev);
719 	if (r)
720 		goto err;
721 
722 	if (index >= VHOST_NET_VQ_MAX) {
723 		r = -ENOBUFS;
724 		goto err;
725 	}
726 	vq = n->vqs + index;
727 	mutex_lock(&vq->mutex);
728 
729 	/* Verify that ring has been setup correctly. */
730 	if (!vhost_vq_access_ok(vq)) {
731 		r = -EFAULT;
732 		goto err_vq;
733 	}
734 	sock = get_socket(fd);
735 	if (IS_ERR(sock)) {
736 		r = PTR_ERR(sock);
737 		goto err_vq;
738 	}
739 
740 	/* start polling new socket */
741 	oldsock = rcu_dereference_protected(vq->private_data,
742 					    lockdep_is_held(&vq->mutex));
743 	if (sock != oldsock) {
744                 vhost_net_disable_vq(n, vq);
745                 rcu_assign_pointer(vq->private_data, sock);
746                 vhost_net_enable_vq(n, vq);
747 	}
748 
749 	mutex_unlock(&vq->mutex);
750 
751 	if (oldsock) {
752 		vhost_net_flush_vq(n, index);
753 		fput(oldsock->file);
754 	}
755 
756 	mutex_unlock(&n->dev.mutex);
757 	return 0;
758 
759 err_vq:
760 	mutex_unlock(&vq->mutex);
761 err:
762 	mutex_unlock(&n->dev.mutex);
763 	return r;
764 }
765 
766 static long vhost_net_reset_owner(struct vhost_net *n)
767 {
768 	struct socket *tx_sock = NULL;
769 	struct socket *rx_sock = NULL;
770 	long err;
771 	mutex_lock(&n->dev.mutex);
772 	err = vhost_dev_check_owner(&n->dev);
773 	if (err)
774 		goto done;
775 	vhost_net_stop(n, &tx_sock, &rx_sock);
776 	vhost_net_flush(n);
777 	err = vhost_dev_reset_owner(&n->dev);
778 done:
779 	mutex_unlock(&n->dev.mutex);
780 	if (tx_sock)
781 		fput(tx_sock->file);
782 	if (rx_sock)
783 		fput(rx_sock->file);
784 	return err;
785 }
786 
787 static int vhost_net_set_features(struct vhost_net *n, u64 features)
788 {
789 	size_t vhost_hlen, sock_hlen, hdr_len;
790 	int i;
791 
792 	hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
793 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
794 			sizeof(struct virtio_net_hdr);
795 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
796 		/* vhost provides vnet_hdr */
797 		vhost_hlen = hdr_len;
798 		sock_hlen = 0;
799 	} else {
800 		/* socket provides vnet_hdr */
801 		vhost_hlen = 0;
802 		sock_hlen = hdr_len;
803 	}
804 	mutex_lock(&n->dev.mutex);
805 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
806 	    !vhost_log_access_ok(&n->dev)) {
807 		mutex_unlock(&n->dev.mutex);
808 		return -EFAULT;
809 	}
810 	n->dev.acked_features = features;
811 	smp_wmb();
812 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
813 		mutex_lock(&n->vqs[i].mutex);
814 		n->vqs[i].vhost_hlen = vhost_hlen;
815 		n->vqs[i].sock_hlen = sock_hlen;
816 		mutex_unlock(&n->vqs[i].mutex);
817 	}
818 	vhost_net_flush(n);
819 	mutex_unlock(&n->dev.mutex);
820 	return 0;
821 }
822 
823 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
824 			    unsigned long arg)
825 {
826 	struct vhost_net *n = f->private_data;
827 	void __user *argp = (void __user *)arg;
828 	u64 __user *featurep = argp;
829 	struct vhost_vring_file backend;
830 	u64 features;
831 	int r;
832 	switch (ioctl) {
833 	case VHOST_NET_SET_BACKEND:
834 		if (copy_from_user(&backend, argp, sizeof backend))
835 			return -EFAULT;
836 		return vhost_net_set_backend(n, backend.index, backend.fd);
837 	case VHOST_GET_FEATURES:
838 		features = VHOST_FEATURES;
839 		if (copy_to_user(featurep, &features, sizeof features))
840 			return -EFAULT;
841 		return 0;
842 	case VHOST_SET_FEATURES:
843 		if (copy_from_user(&features, featurep, sizeof features))
844 			return -EFAULT;
845 		if (features & ~VHOST_FEATURES)
846 			return -EOPNOTSUPP;
847 		return vhost_net_set_features(n, features);
848 	case VHOST_RESET_OWNER:
849 		return vhost_net_reset_owner(n);
850 	default:
851 		mutex_lock(&n->dev.mutex);
852 		r = vhost_dev_ioctl(&n->dev, ioctl, arg);
853 		vhost_net_flush(n);
854 		mutex_unlock(&n->dev.mutex);
855 		return r;
856 	}
857 }
858 
859 #ifdef CONFIG_COMPAT
860 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
861 				   unsigned long arg)
862 {
863 	return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
864 }
865 #endif
866 
867 static const struct file_operations vhost_net_fops = {
868 	.owner          = THIS_MODULE,
869 	.release        = vhost_net_release,
870 	.unlocked_ioctl = vhost_net_ioctl,
871 #ifdef CONFIG_COMPAT
872 	.compat_ioctl   = vhost_net_compat_ioctl,
873 #endif
874 	.open           = vhost_net_open,
875 	.llseek		= noop_llseek,
876 };
877 
878 static struct miscdevice vhost_net_misc = {
879 	MISC_DYNAMIC_MINOR,
880 	"vhost-net",
881 	&vhost_net_fops,
882 };
883 
884 static int vhost_net_init(void)
885 {
886 	return misc_register(&vhost_net_misc);
887 }
888 module_init(vhost_net_init);
889 
890 static void vhost_net_exit(void)
891 {
892 	misc_deregister(&vhost_net_misc);
893 }
894 module_exit(vhost_net_exit);
895 
896 MODULE_VERSION("0.0.1");
897 MODULE_LICENSE("GPL v2");
898 MODULE_AUTHOR("Michael S. Tsirkin");
899 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
900