xref: /linux/drivers/vhost/net.c (revision b889fcf63cb62e7fdb7816565e28f44dbe4a76a5)
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/moduleparam.h>
16 #include <linux/mutex.h>
17 #include <linux/workqueue.h>
18 #include <linux/rcupdate.h>
19 #include <linux/file.h>
20 #include <linux/slab.h>
21 
22 #include <linux/net.h>
23 #include <linux/if_packet.h>
24 #include <linux/if_arp.h>
25 #include <linux/if_tun.h>
26 #include <linux/if_macvlan.h>
27 #include <linux/if_vlan.h>
28 
29 #include <net/sock.h>
30 
31 #include "vhost.h"
32 
33 static int experimental_zcopytx = 1;
34 module_param(experimental_zcopytx, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
36 		                       " 1 -Enable; 0 - Disable");
37 
38 /* Max number of bytes transferred before requeueing the job.
39  * Using this limit prevents one virtqueue from starving others. */
40 #define VHOST_NET_WEIGHT 0x80000
41 
42 /* MAX number of TX used buffers for outstanding zerocopy */
43 #define VHOST_MAX_PEND 128
44 #define VHOST_GOODCOPY_LEN 256
45 
46 /*
47  * For transmit, used buffer len is unused; we override it to track buffer
48  * status internally; used for zerocopy tx only.
49  */
50 /* Lower device DMA failed */
51 #define VHOST_DMA_FAILED_LEN	3
52 /* Lower device DMA done */
53 #define VHOST_DMA_DONE_LEN	2
54 /* Lower device DMA in progress */
55 #define VHOST_DMA_IN_PROGRESS	1
56 /* Buffer unused */
57 #define VHOST_DMA_CLEAR_LEN	0
58 
59 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
60 
61 enum {
62 	VHOST_NET_VQ_RX = 0,
63 	VHOST_NET_VQ_TX = 1,
64 	VHOST_NET_VQ_MAX = 2,
65 };
66 
67 enum vhost_net_poll_state {
68 	VHOST_NET_POLL_DISABLED = 0,
69 	VHOST_NET_POLL_STARTED = 1,
70 	VHOST_NET_POLL_STOPPED = 2,
71 };
72 
73 struct vhost_net {
74 	struct vhost_dev dev;
75 	struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
76 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
77 	/* Tells us whether we are polling a socket for TX.
78 	 * We only do this when socket buffer fills up.
79 	 * Protected by tx vq lock. */
80 	enum vhost_net_poll_state tx_poll_state;
81 	/* Number of TX recently submitted.
82 	 * Protected by tx vq lock. */
83 	unsigned tx_packets;
84 	/* Number of times zerocopy TX recently failed.
85 	 * Protected by tx vq lock. */
86 	unsigned tx_zcopy_err;
87 	/* Flush in progress. Protected by tx vq lock. */
88 	bool tx_flush;
89 };
90 
91 static void vhost_net_tx_packet(struct vhost_net *net)
92 {
93 	++net->tx_packets;
94 	if (net->tx_packets < 1024)
95 		return;
96 	net->tx_packets = 0;
97 	net->tx_zcopy_err = 0;
98 }
99 
100 static void vhost_net_tx_err(struct vhost_net *net)
101 {
102 	++net->tx_zcopy_err;
103 }
104 
105 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
106 {
107 	/* TX flush waits for outstanding DMAs to be done.
108 	 * Don't start new DMAs.
109 	 */
110 	return !net->tx_flush &&
111 		net->tx_packets / 64 >= net->tx_zcopy_err;
112 }
113 
114 static bool vhost_sock_zcopy(struct socket *sock)
115 {
116 	return unlikely(experimental_zcopytx) &&
117 		sock_flag(sock->sk, SOCK_ZEROCOPY);
118 }
119 
120 /* Pop first len bytes from iovec. Return number of segments used. */
121 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
122 			  size_t len, int iov_count)
123 {
124 	int seg = 0;
125 	size_t size;
126 
127 	while (len && seg < iov_count) {
128 		size = min(from->iov_len, len);
129 		to->iov_base = from->iov_base;
130 		to->iov_len = size;
131 		from->iov_len -= size;
132 		from->iov_base += size;
133 		len -= size;
134 		++from;
135 		++to;
136 		++seg;
137 	}
138 	return seg;
139 }
140 /* Copy iovec entries for len bytes from iovec. */
141 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
142 			   size_t len, int iovcount)
143 {
144 	int seg = 0;
145 	size_t size;
146 
147 	while (len && seg < iovcount) {
148 		size = min(from->iov_len, len);
149 		to->iov_base = from->iov_base;
150 		to->iov_len = size;
151 		len -= size;
152 		++from;
153 		++to;
154 		++seg;
155 	}
156 }
157 
158 /* Caller must have TX VQ lock */
159 static void tx_poll_stop(struct vhost_net *net)
160 {
161 	if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
162 		return;
163 	vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
164 	net->tx_poll_state = VHOST_NET_POLL_STOPPED;
165 }
166 
167 /* Caller must have TX VQ lock */
168 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
169 {
170 	if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
171 		return;
172 	vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
173 	net->tx_poll_state = VHOST_NET_POLL_STARTED;
174 }
175 
176 /* In case of DMA done not in order in lower device driver for some reason.
177  * upend_idx is used to track end of used idx, done_idx is used to track head
178  * of used idx. Once lower device DMA done contiguously, we will signal KVM
179  * guest used idx.
180  */
181 static int vhost_zerocopy_signal_used(struct vhost_net *net,
182 				      struct vhost_virtqueue *vq)
183 {
184 	int i;
185 	int j = 0;
186 
187 	for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
188 		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
189 			vhost_net_tx_err(net);
190 		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
191 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
192 			vhost_add_used_and_signal(vq->dev, vq,
193 						  vq->heads[i].id, 0);
194 			++j;
195 		} else
196 			break;
197 	}
198 	if (j)
199 		vq->done_idx = i;
200 	return j;
201 }
202 
203 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
204 {
205 	struct vhost_ubuf_ref *ubufs = ubuf->ctx;
206 	struct vhost_virtqueue *vq = ubufs->vq;
207 	int cnt = atomic_read(&ubufs->kref.refcount);
208 
209 	/*
210 	 * Trigger polling thread if guest stopped submitting new buffers:
211 	 * in this case, the refcount after decrement will eventually reach 1
212 	 * so here it is 2.
213 	 * We also trigger polling periodically after each 16 packets
214 	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
215 	 * less than 10% of times).
216 	 */
217 	if (cnt <= 2 || !(cnt % 16))
218 		vhost_poll_queue(&vq->poll);
219 	/* set len to mark this desc buffers done DMA */
220 	vq->heads[ubuf->desc].len = success ?
221 		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
222 	vhost_ubuf_put(ubufs);
223 }
224 
225 /* Expects to be always run from workqueue - which acts as
226  * read-size critical section for our kind of RCU. */
227 static void handle_tx(struct vhost_net *net)
228 {
229 	struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
230 	unsigned out, in, s;
231 	int head;
232 	struct msghdr msg = {
233 		.msg_name = NULL,
234 		.msg_namelen = 0,
235 		.msg_control = NULL,
236 		.msg_controllen = 0,
237 		.msg_iov = vq->iov,
238 		.msg_flags = MSG_DONTWAIT,
239 	};
240 	size_t len, total_len = 0;
241 	int err, wmem;
242 	size_t hdr_size;
243 	struct socket *sock;
244 	struct vhost_ubuf_ref *uninitialized_var(ubufs);
245 	bool zcopy, zcopy_used;
246 
247 	/* TODO: check that we are running from vhost_worker? */
248 	sock = rcu_dereference_check(vq->private_data, 1);
249 	if (!sock)
250 		return;
251 
252 	wmem = atomic_read(&sock->sk->sk_wmem_alloc);
253 	if (wmem >= sock->sk->sk_sndbuf) {
254 		mutex_lock(&vq->mutex);
255 		tx_poll_start(net, sock);
256 		mutex_unlock(&vq->mutex);
257 		return;
258 	}
259 
260 	mutex_lock(&vq->mutex);
261 	vhost_disable_notify(&net->dev, vq);
262 
263 	if (wmem < sock->sk->sk_sndbuf / 2)
264 		tx_poll_stop(net);
265 	hdr_size = vq->vhost_hlen;
266 	zcopy = vq->ubufs;
267 
268 	for (;;) {
269 		/* Release DMAs done buffers first */
270 		if (zcopy)
271 			vhost_zerocopy_signal_used(net, vq);
272 
273 		head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
274 					 ARRAY_SIZE(vq->iov),
275 					 &out, &in,
276 					 NULL, NULL);
277 		/* On error, stop handling until the next kick. */
278 		if (unlikely(head < 0))
279 			break;
280 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
281 		if (head == vq->num) {
282 			int num_pends;
283 
284 			wmem = atomic_read(&sock->sk->sk_wmem_alloc);
285 			if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
286 				tx_poll_start(net, sock);
287 				set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
288 				break;
289 			}
290 			/* If more outstanding DMAs, queue the work.
291 			 * Handle upend_idx wrap around
292 			 */
293 			num_pends = likely(vq->upend_idx >= vq->done_idx) ?
294 				    (vq->upend_idx - vq->done_idx) :
295 				    (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
296 			if (unlikely(num_pends > VHOST_MAX_PEND)) {
297 				tx_poll_start(net, sock);
298 				set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
299 				break;
300 			}
301 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
302 				vhost_disable_notify(&net->dev, vq);
303 				continue;
304 			}
305 			break;
306 		}
307 		if (in) {
308 			vq_err(vq, "Unexpected descriptor format for TX: "
309 			       "out %d, int %d\n", out, in);
310 			break;
311 		}
312 		/* Skip header. TODO: support TSO. */
313 		s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
314 		msg.msg_iovlen = out;
315 		len = iov_length(vq->iov, out);
316 		/* Sanity check */
317 		if (!len) {
318 			vq_err(vq, "Unexpected header len for TX: "
319 			       "%zd expected %zd\n",
320 			       iov_length(vq->hdr, s), hdr_size);
321 			break;
322 		}
323 		zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
324 				       vq->upend_idx != vq->done_idx);
325 
326 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
327 		if (zcopy_used) {
328 			vq->heads[vq->upend_idx].id = head;
329 			if (!vhost_net_tx_select_zcopy(net) ||
330 			    len < VHOST_GOODCOPY_LEN) {
331 				/* copy don't need to wait for DMA done */
332 				vq->heads[vq->upend_idx].len =
333 							VHOST_DMA_DONE_LEN;
334 				msg.msg_control = NULL;
335 				msg.msg_controllen = 0;
336 				ubufs = NULL;
337 			} else {
338 				struct ubuf_info *ubuf = &vq->ubuf_info[head];
339 
340 				vq->heads[vq->upend_idx].len =
341 					VHOST_DMA_IN_PROGRESS;
342 				ubuf->callback = vhost_zerocopy_callback;
343 				ubuf->ctx = vq->ubufs;
344 				ubuf->desc = vq->upend_idx;
345 				msg.msg_control = ubuf;
346 				msg.msg_controllen = sizeof(ubuf);
347 				ubufs = vq->ubufs;
348 				kref_get(&ubufs->kref);
349 			}
350 			vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
351 		}
352 		/* TODO: Check specific error and bomb out unless ENOBUFS? */
353 		err = sock->ops->sendmsg(NULL, sock, &msg, len);
354 		if (unlikely(err < 0)) {
355 			if (zcopy_used) {
356 				if (ubufs)
357 					vhost_ubuf_put(ubufs);
358 				vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
359 					UIO_MAXIOV;
360 			}
361 			vhost_discard_vq_desc(vq, 1);
362 			if (err == -EAGAIN || err == -ENOBUFS)
363 				tx_poll_start(net, sock);
364 			break;
365 		}
366 		if (err != len)
367 			pr_debug("Truncated TX packet: "
368 				 " len %d != %zd\n", err, len);
369 		if (!zcopy_used)
370 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
371 		else
372 			vhost_zerocopy_signal_used(net, vq);
373 		total_len += len;
374 		vhost_net_tx_packet(net);
375 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
376 			vhost_poll_queue(&vq->poll);
377 			break;
378 		}
379 	}
380 
381 	mutex_unlock(&vq->mutex);
382 }
383 
384 static int peek_head_len(struct sock *sk)
385 {
386 	struct sk_buff *head;
387 	int len = 0;
388 	unsigned long flags;
389 
390 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
391 	head = skb_peek(&sk->sk_receive_queue);
392 	if (likely(head)) {
393 		len = head->len;
394 		if (vlan_tx_tag_present(head))
395 			len += VLAN_HLEN;
396 	}
397 
398 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
399 	return len;
400 }
401 
402 /* This is a multi-buffer version of vhost_get_desc, that works if
403  *	vq has read descriptors only.
404  * @vq		- the relevant virtqueue
405  * @datalen	- data length we'll be reading
406  * @iovcount	- returned count of io vectors we fill
407  * @log		- vhost log
408  * @log_num	- log offset
409  * @quota       - headcount quota, 1 for big buffer
410  *	returns number of buffer heads allocated, negative on error
411  */
412 static int get_rx_bufs(struct vhost_virtqueue *vq,
413 		       struct vring_used_elem *heads,
414 		       int datalen,
415 		       unsigned *iovcount,
416 		       struct vhost_log *log,
417 		       unsigned *log_num,
418 		       unsigned int quota)
419 {
420 	unsigned int out, in;
421 	int seg = 0;
422 	int headcount = 0;
423 	unsigned d;
424 	int r, nlogs = 0;
425 
426 	while (datalen > 0 && headcount < quota) {
427 		if (unlikely(seg >= UIO_MAXIOV)) {
428 			r = -ENOBUFS;
429 			goto err;
430 		}
431 		d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
432 				      ARRAY_SIZE(vq->iov) - seg, &out,
433 				      &in, log, log_num);
434 		if (d == vq->num) {
435 			r = 0;
436 			goto err;
437 		}
438 		if (unlikely(out || in <= 0)) {
439 			vq_err(vq, "unexpected descriptor format for RX: "
440 				"out %d, in %d\n", out, in);
441 			r = -EINVAL;
442 			goto err;
443 		}
444 		if (unlikely(log)) {
445 			nlogs += *log_num;
446 			log += *log_num;
447 		}
448 		heads[headcount].id = d;
449 		heads[headcount].len = iov_length(vq->iov + seg, in);
450 		datalen -= heads[headcount].len;
451 		++headcount;
452 		seg += in;
453 	}
454 	heads[headcount - 1].len += datalen;
455 	*iovcount = seg;
456 	if (unlikely(log))
457 		*log_num = nlogs;
458 	return headcount;
459 err:
460 	vhost_discard_vq_desc(vq, headcount);
461 	return r;
462 }
463 
464 /* Expects to be always run from workqueue - which acts as
465  * read-size critical section for our kind of RCU. */
466 static void handle_rx(struct vhost_net *net)
467 {
468 	struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
469 	unsigned uninitialized_var(in), log;
470 	struct vhost_log *vq_log;
471 	struct msghdr msg = {
472 		.msg_name = NULL,
473 		.msg_namelen = 0,
474 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
475 		.msg_controllen = 0,
476 		.msg_iov = vq->iov,
477 		.msg_flags = MSG_DONTWAIT,
478 	};
479 	struct virtio_net_hdr_mrg_rxbuf hdr = {
480 		.hdr.flags = 0,
481 		.hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
482 	};
483 	size_t total_len = 0;
484 	int err, mergeable;
485 	s16 headcount;
486 	size_t vhost_hlen, sock_hlen;
487 	size_t vhost_len, sock_len;
488 	/* TODO: check that we are running from vhost_worker? */
489 	struct socket *sock = rcu_dereference_check(vq->private_data, 1);
490 
491 	if (!sock)
492 		return;
493 
494 	mutex_lock(&vq->mutex);
495 	vhost_disable_notify(&net->dev, vq);
496 	vhost_hlen = vq->vhost_hlen;
497 	sock_hlen = vq->sock_hlen;
498 
499 	vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
500 		vq->log : NULL;
501 	mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
502 
503 	while ((sock_len = peek_head_len(sock->sk))) {
504 		sock_len += sock_hlen;
505 		vhost_len = sock_len + vhost_hlen;
506 		headcount = get_rx_bufs(vq, vq->heads, vhost_len,
507 					&in, vq_log, &log,
508 					likely(mergeable) ? UIO_MAXIOV : 1);
509 		/* On error, stop handling until the next kick. */
510 		if (unlikely(headcount < 0))
511 			break;
512 		/* OK, now we need to know about added descriptors. */
513 		if (!headcount) {
514 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
515 				/* They have slipped one in as we were
516 				 * doing that: check again. */
517 				vhost_disable_notify(&net->dev, vq);
518 				continue;
519 			}
520 			/* Nothing new?  Wait for eventfd to tell us
521 			 * they refilled. */
522 			break;
523 		}
524 		/* We don't need to be notified again. */
525 		if (unlikely((vhost_hlen)))
526 			/* Skip header. TODO: support TSO. */
527 			move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
528 		else
529 			/* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
530 			 * needed because recvmsg can modify msg_iov. */
531 			copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
532 		msg.msg_iovlen = in;
533 		err = sock->ops->recvmsg(NULL, sock, &msg,
534 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
535 		/* Userspace might have consumed the packet meanwhile:
536 		 * it's not supposed to do this usually, but might be hard
537 		 * to prevent. Discard data we got (if any) and keep going. */
538 		if (unlikely(err != sock_len)) {
539 			pr_debug("Discarded rx packet: "
540 				 " len %d, expected %zd\n", err, sock_len);
541 			vhost_discard_vq_desc(vq, headcount);
542 			continue;
543 		}
544 		if (unlikely(vhost_hlen) &&
545 		    memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
546 				      vhost_hlen)) {
547 			vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
548 			       vq->iov->iov_base);
549 			break;
550 		}
551 		/* TODO: Should check and handle checksum. */
552 		if (likely(mergeable) &&
553 		    memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
554 				      offsetof(typeof(hdr), num_buffers),
555 				      sizeof hdr.num_buffers)) {
556 			vq_err(vq, "Failed num_buffers write");
557 			vhost_discard_vq_desc(vq, headcount);
558 			break;
559 		}
560 		vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
561 					    headcount);
562 		if (unlikely(vq_log))
563 			vhost_log_write(vq, vq_log, log, vhost_len);
564 		total_len += vhost_len;
565 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
566 			vhost_poll_queue(&vq->poll);
567 			break;
568 		}
569 	}
570 
571 	mutex_unlock(&vq->mutex);
572 }
573 
574 static void handle_tx_kick(struct vhost_work *work)
575 {
576 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
577 						  poll.work);
578 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
579 
580 	handle_tx(net);
581 }
582 
583 static void handle_rx_kick(struct vhost_work *work)
584 {
585 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
586 						  poll.work);
587 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
588 
589 	handle_rx(net);
590 }
591 
592 static void handle_tx_net(struct vhost_work *work)
593 {
594 	struct vhost_net *net = container_of(work, struct vhost_net,
595 					     poll[VHOST_NET_VQ_TX].work);
596 	handle_tx(net);
597 }
598 
599 static void handle_rx_net(struct vhost_work *work)
600 {
601 	struct vhost_net *net = container_of(work, struct vhost_net,
602 					     poll[VHOST_NET_VQ_RX].work);
603 	handle_rx(net);
604 }
605 
606 static int vhost_net_open(struct inode *inode, struct file *f)
607 {
608 	struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
609 	struct vhost_dev *dev;
610 	int r;
611 
612 	if (!n)
613 		return -ENOMEM;
614 
615 	dev = &n->dev;
616 	n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
617 	n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
618 	r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
619 	if (r < 0) {
620 		kfree(n);
621 		return r;
622 	}
623 
624 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
625 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
626 	n->tx_poll_state = VHOST_NET_POLL_DISABLED;
627 
628 	f->private_data = n;
629 
630 	return 0;
631 }
632 
633 static void vhost_net_disable_vq(struct vhost_net *n,
634 				 struct vhost_virtqueue *vq)
635 {
636 	if (!vq->private_data)
637 		return;
638 	if (vq == n->vqs + VHOST_NET_VQ_TX) {
639 		tx_poll_stop(n);
640 		n->tx_poll_state = VHOST_NET_POLL_DISABLED;
641 	} else
642 		vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
643 }
644 
645 static void vhost_net_enable_vq(struct vhost_net *n,
646 				struct vhost_virtqueue *vq)
647 {
648 	struct socket *sock;
649 
650 	sock = rcu_dereference_protected(vq->private_data,
651 					 lockdep_is_held(&vq->mutex));
652 	if (!sock)
653 		return;
654 	if (vq == n->vqs + VHOST_NET_VQ_TX) {
655 		n->tx_poll_state = VHOST_NET_POLL_STOPPED;
656 		tx_poll_start(n, sock);
657 	} else
658 		vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
659 }
660 
661 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
662 					struct vhost_virtqueue *vq)
663 {
664 	struct socket *sock;
665 
666 	mutex_lock(&vq->mutex);
667 	sock = rcu_dereference_protected(vq->private_data,
668 					 lockdep_is_held(&vq->mutex));
669 	vhost_net_disable_vq(n, vq);
670 	rcu_assign_pointer(vq->private_data, NULL);
671 	mutex_unlock(&vq->mutex);
672 	return sock;
673 }
674 
675 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
676 			   struct socket **rx_sock)
677 {
678 	*tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
679 	*rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
680 }
681 
682 static void vhost_net_flush_vq(struct vhost_net *n, int index)
683 {
684 	vhost_poll_flush(n->poll + index);
685 	vhost_poll_flush(&n->dev.vqs[index].poll);
686 }
687 
688 static void vhost_net_flush(struct vhost_net *n)
689 {
690 	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
691 	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
692 	if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
693 		mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
694 		n->tx_flush = true;
695 		mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
696 		/* Wait for all lower device DMAs done. */
697 		vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
698 		mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
699 		n->tx_flush = false;
700 		kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
701 		mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
702 	}
703 }
704 
705 static int vhost_net_release(struct inode *inode, struct file *f)
706 {
707 	struct vhost_net *n = f->private_data;
708 	struct socket *tx_sock;
709 	struct socket *rx_sock;
710 
711 	vhost_net_stop(n, &tx_sock, &rx_sock);
712 	vhost_net_flush(n);
713 	vhost_dev_stop(&n->dev);
714 	vhost_dev_cleanup(&n->dev, false);
715 	if (tx_sock)
716 		fput(tx_sock->file);
717 	if (rx_sock)
718 		fput(rx_sock->file);
719 	/* We do an extra flush before freeing memory,
720 	 * since jobs can re-queue themselves. */
721 	vhost_net_flush(n);
722 	kfree(n);
723 	return 0;
724 }
725 
726 static struct socket *get_raw_socket(int fd)
727 {
728 	struct {
729 		struct sockaddr_ll sa;
730 		char  buf[MAX_ADDR_LEN];
731 	} uaddr;
732 	int uaddr_len = sizeof uaddr, r;
733 	struct socket *sock = sockfd_lookup(fd, &r);
734 
735 	if (!sock)
736 		return ERR_PTR(-ENOTSOCK);
737 
738 	/* Parameter checking */
739 	if (sock->sk->sk_type != SOCK_RAW) {
740 		r = -ESOCKTNOSUPPORT;
741 		goto err;
742 	}
743 
744 	r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
745 			       &uaddr_len, 0);
746 	if (r)
747 		goto err;
748 
749 	if (uaddr.sa.sll_family != AF_PACKET) {
750 		r = -EPFNOSUPPORT;
751 		goto err;
752 	}
753 	return sock;
754 err:
755 	fput(sock->file);
756 	return ERR_PTR(r);
757 }
758 
759 static struct socket *get_tap_socket(int fd)
760 {
761 	struct file *file = fget(fd);
762 	struct socket *sock;
763 
764 	if (!file)
765 		return ERR_PTR(-EBADF);
766 	sock = tun_get_socket(file);
767 	if (!IS_ERR(sock))
768 		return sock;
769 	sock = macvtap_get_socket(file);
770 	if (IS_ERR(sock))
771 		fput(file);
772 	return sock;
773 }
774 
775 static struct socket *get_socket(int fd)
776 {
777 	struct socket *sock;
778 
779 	/* special case to disable backend */
780 	if (fd == -1)
781 		return NULL;
782 	sock = get_raw_socket(fd);
783 	if (!IS_ERR(sock))
784 		return sock;
785 	sock = get_tap_socket(fd);
786 	if (!IS_ERR(sock))
787 		return sock;
788 	return ERR_PTR(-ENOTSOCK);
789 }
790 
791 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
792 {
793 	struct socket *sock, *oldsock;
794 	struct vhost_virtqueue *vq;
795 	struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
796 	int r;
797 
798 	mutex_lock(&n->dev.mutex);
799 	r = vhost_dev_check_owner(&n->dev);
800 	if (r)
801 		goto err;
802 
803 	if (index >= VHOST_NET_VQ_MAX) {
804 		r = -ENOBUFS;
805 		goto err;
806 	}
807 	vq = n->vqs + index;
808 	mutex_lock(&vq->mutex);
809 
810 	/* Verify that ring has been setup correctly. */
811 	if (!vhost_vq_access_ok(vq)) {
812 		r = -EFAULT;
813 		goto err_vq;
814 	}
815 	sock = get_socket(fd);
816 	if (IS_ERR(sock)) {
817 		r = PTR_ERR(sock);
818 		goto err_vq;
819 	}
820 
821 	/* start polling new socket */
822 	oldsock = rcu_dereference_protected(vq->private_data,
823 					    lockdep_is_held(&vq->mutex));
824 	if (sock != oldsock) {
825 		ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
826 		if (IS_ERR(ubufs)) {
827 			r = PTR_ERR(ubufs);
828 			goto err_ubufs;
829 		}
830 		oldubufs = vq->ubufs;
831 		vq->ubufs = ubufs;
832 		vhost_net_disable_vq(n, vq);
833 		rcu_assign_pointer(vq->private_data, sock);
834 		vhost_net_enable_vq(n, vq);
835 
836 		r = vhost_init_used(vq);
837 		if (r)
838 			goto err_vq;
839 
840 		n->tx_packets = 0;
841 		n->tx_zcopy_err = 0;
842 		n->tx_flush = false;
843 	}
844 
845 	mutex_unlock(&vq->mutex);
846 
847 	if (oldubufs) {
848 		vhost_ubuf_put_and_wait(oldubufs);
849 		mutex_lock(&vq->mutex);
850 		vhost_zerocopy_signal_used(n, vq);
851 		mutex_unlock(&vq->mutex);
852 	}
853 
854 	if (oldsock) {
855 		vhost_net_flush_vq(n, index);
856 		fput(oldsock->file);
857 	}
858 
859 	mutex_unlock(&n->dev.mutex);
860 	return 0;
861 
862 err_ubufs:
863 	fput(sock->file);
864 err_vq:
865 	mutex_unlock(&vq->mutex);
866 err:
867 	mutex_unlock(&n->dev.mutex);
868 	return r;
869 }
870 
871 static long vhost_net_reset_owner(struct vhost_net *n)
872 {
873 	struct socket *tx_sock = NULL;
874 	struct socket *rx_sock = NULL;
875 	long err;
876 
877 	mutex_lock(&n->dev.mutex);
878 	err = vhost_dev_check_owner(&n->dev);
879 	if (err)
880 		goto done;
881 	vhost_net_stop(n, &tx_sock, &rx_sock);
882 	vhost_net_flush(n);
883 	err = vhost_dev_reset_owner(&n->dev);
884 done:
885 	mutex_unlock(&n->dev.mutex);
886 	if (tx_sock)
887 		fput(tx_sock->file);
888 	if (rx_sock)
889 		fput(rx_sock->file);
890 	return err;
891 }
892 
893 static int vhost_net_set_features(struct vhost_net *n, u64 features)
894 {
895 	size_t vhost_hlen, sock_hlen, hdr_len;
896 	int i;
897 
898 	hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
899 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
900 			sizeof(struct virtio_net_hdr);
901 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
902 		/* vhost provides vnet_hdr */
903 		vhost_hlen = hdr_len;
904 		sock_hlen = 0;
905 	} else {
906 		/* socket provides vnet_hdr */
907 		vhost_hlen = 0;
908 		sock_hlen = hdr_len;
909 	}
910 	mutex_lock(&n->dev.mutex);
911 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
912 	    !vhost_log_access_ok(&n->dev)) {
913 		mutex_unlock(&n->dev.mutex);
914 		return -EFAULT;
915 	}
916 	n->dev.acked_features = features;
917 	smp_wmb();
918 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
919 		mutex_lock(&n->vqs[i].mutex);
920 		n->vqs[i].vhost_hlen = vhost_hlen;
921 		n->vqs[i].sock_hlen = sock_hlen;
922 		mutex_unlock(&n->vqs[i].mutex);
923 	}
924 	vhost_net_flush(n);
925 	mutex_unlock(&n->dev.mutex);
926 	return 0;
927 }
928 
929 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
930 			    unsigned long arg)
931 {
932 	struct vhost_net *n = f->private_data;
933 	void __user *argp = (void __user *)arg;
934 	u64 __user *featurep = argp;
935 	struct vhost_vring_file backend;
936 	u64 features;
937 	int r;
938 
939 	switch (ioctl) {
940 	case VHOST_NET_SET_BACKEND:
941 		if (copy_from_user(&backend, argp, sizeof backend))
942 			return -EFAULT;
943 		return vhost_net_set_backend(n, backend.index, backend.fd);
944 	case VHOST_GET_FEATURES:
945 		features = VHOST_NET_FEATURES;
946 		if (copy_to_user(featurep, &features, sizeof features))
947 			return -EFAULT;
948 		return 0;
949 	case VHOST_SET_FEATURES:
950 		if (copy_from_user(&features, featurep, sizeof features))
951 			return -EFAULT;
952 		if (features & ~VHOST_NET_FEATURES)
953 			return -EOPNOTSUPP;
954 		return vhost_net_set_features(n, features);
955 	case VHOST_RESET_OWNER:
956 		return vhost_net_reset_owner(n);
957 	default:
958 		mutex_lock(&n->dev.mutex);
959 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
960 		if (r == -ENOIOCTLCMD)
961 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
962 		else
963 			vhost_net_flush(n);
964 		mutex_unlock(&n->dev.mutex);
965 		return r;
966 	}
967 }
968 
969 #ifdef CONFIG_COMPAT
970 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
971 				   unsigned long arg)
972 {
973 	return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
974 }
975 #endif
976 
977 static const struct file_operations vhost_net_fops = {
978 	.owner          = THIS_MODULE,
979 	.release        = vhost_net_release,
980 	.unlocked_ioctl = vhost_net_ioctl,
981 #ifdef CONFIG_COMPAT
982 	.compat_ioctl   = vhost_net_compat_ioctl,
983 #endif
984 	.open           = vhost_net_open,
985 	.llseek		= noop_llseek,
986 };
987 
988 static struct miscdevice vhost_net_misc = {
989 	.minor = VHOST_NET_MINOR,
990 	.name = "vhost-net",
991 	.fops = &vhost_net_fops,
992 };
993 
994 static int vhost_net_init(void)
995 {
996 	if (experimental_zcopytx)
997 		vhost_enable_zcopy(VHOST_NET_VQ_TX);
998 	return misc_register(&vhost_net_misc);
999 }
1000 module_init(vhost_net_init);
1001 
1002 static void vhost_net_exit(void)
1003 {
1004 	misc_deregister(&vhost_net_misc);
1005 }
1006 module_exit(vhost_net_exit);
1007 
1008 MODULE_VERSION("0.0.1");
1009 MODULE_LICENSE("GPL v2");
1010 MODULE_AUTHOR("Michael S. Tsirkin");
1011 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1012 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1013 MODULE_ALIAS("devname:vhost-net");
1014