xref: /linux/drivers/vhost/net.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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/file.h>
19 #include <linux/slab.h>
20 #include <linux/vmalloc.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	((__force __virtio32)3)
52 /* Lower device DMA done */
53 #define VHOST_DMA_DONE_LEN	((__force __virtio32)2)
54 /* Lower device DMA in progress */
55 #define VHOST_DMA_IN_PROGRESS	((__force __virtio32)1)
56 /* Buffer unused */
57 #define VHOST_DMA_CLEAR_LEN	((__force __virtio32)0)
58 
59 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
60 
61 enum {
62 	VHOST_NET_FEATURES = VHOST_FEATURES |
63 			 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
64 			 (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
65 			 (1ULL << VIRTIO_F_VERSION_1),
66 };
67 
68 enum {
69 	VHOST_NET_VQ_RX = 0,
70 	VHOST_NET_VQ_TX = 1,
71 	VHOST_NET_VQ_MAX = 2,
72 };
73 
74 struct vhost_net_ubuf_ref {
75 	/* refcount follows semantics similar to kref:
76 	 *  0: object is released
77 	 *  1: no outstanding ubufs
78 	 * >1: outstanding ubufs
79 	 */
80 	atomic_t refcount;
81 	wait_queue_head_t wait;
82 	struct vhost_virtqueue *vq;
83 };
84 
85 struct vhost_net_virtqueue {
86 	struct vhost_virtqueue vq;
87 	size_t vhost_hlen;
88 	size_t sock_hlen;
89 	/* vhost zerocopy support fields below: */
90 	/* last used idx for outstanding DMA zerocopy buffers */
91 	int upend_idx;
92 	/* first used idx for DMA done zerocopy buffers */
93 	int done_idx;
94 	/* an array of userspace buffers info */
95 	struct ubuf_info *ubuf_info;
96 	/* Reference counting for outstanding ubufs.
97 	 * Protected by vq mutex. Writers must also take device mutex. */
98 	struct vhost_net_ubuf_ref *ubufs;
99 };
100 
101 struct vhost_net {
102 	struct vhost_dev dev;
103 	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
104 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
105 	/* Number of TX recently submitted.
106 	 * Protected by tx vq lock. */
107 	unsigned tx_packets;
108 	/* Number of times zerocopy TX recently failed.
109 	 * Protected by tx vq lock. */
110 	unsigned tx_zcopy_err;
111 	/* Flush in progress. Protected by tx vq lock. */
112 	bool tx_flush;
113 };
114 
115 static unsigned vhost_net_zcopy_mask __read_mostly;
116 
117 static void vhost_net_enable_zcopy(int vq)
118 {
119 	vhost_net_zcopy_mask |= 0x1 << vq;
120 }
121 
122 static struct vhost_net_ubuf_ref *
123 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
124 {
125 	struct vhost_net_ubuf_ref *ubufs;
126 	/* No zero copy backend? Nothing to count. */
127 	if (!zcopy)
128 		return NULL;
129 	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
130 	if (!ubufs)
131 		return ERR_PTR(-ENOMEM);
132 	atomic_set(&ubufs->refcount, 1);
133 	init_waitqueue_head(&ubufs->wait);
134 	ubufs->vq = vq;
135 	return ubufs;
136 }
137 
138 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
139 {
140 	int r = atomic_sub_return(1, &ubufs->refcount);
141 	if (unlikely(!r))
142 		wake_up(&ubufs->wait);
143 	return r;
144 }
145 
146 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
147 {
148 	vhost_net_ubuf_put(ubufs);
149 	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
150 }
151 
152 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
153 {
154 	vhost_net_ubuf_put_and_wait(ubufs);
155 	kfree(ubufs);
156 }
157 
158 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
159 {
160 	int i;
161 
162 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
163 		kfree(n->vqs[i].ubuf_info);
164 		n->vqs[i].ubuf_info = NULL;
165 	}
166 }
167 
168 static int vhost_net_set_ubuf_info(struct vhost_net *n)
169 {
170 	bool zcopy;
171 	int i;
172 
173 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
174 		zcopy = vhost_net_zcopy_mask & (0x1 << i);
175 		if (!zcopy)
176 			continue;
177 		n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
178 					      UIO_MAXIOV, GFP_KERNEL);
179 		if  (!n->vqs[i].ubuf_info)
180 			goto err;
181 	}
182 	return 0;
183 
184 err:
185 	vhost_net_clear_ubuf_info(n);
186 	return -ENOMEM;
187 }
188 
189 static void vhost_net_vq_reset(struct vhost_net *n)
190 {
191 	int i;
192 
193 	vhost_net_clear_ubuf_info(n);
194 
195 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
196 		n->vqs[i].done_idx = 0;
197 		n->vqs[i].upend_idx = 0;
198 		n->vqs[i].ubufs = NULL;
199 		n->vqs[i].vhost_hlen = 0;
200 		n->vqs[i].sock_hlen = 0;
201 	}
202 
203 }
204 
205 static void vhost_net_tx_packet(struct vhost_net *net)
206 {
207 	++net->tx_packets;
208 	if (net->tx_packets < 1024)
209 		return;
210 	net->tx_packets = 0;
211 	net->tx_zcopy_err = 0;
212 }
213 
214 static void vhost_net_tx_err(struct vhost_net *net)
215 {
216 	++net->tx_zcopy_err;
217 }
218 
219 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
220 {
221 	/* TX flush waits for outstanding DMAs to be done.
222 	 * Don't start new DMAs.
223 	 */
224 	return !net->tx_flush &&
225 		net->tx_packets / 64 >= net->tx_zcopy_err;
226 }
227 
228 static bool vhost_sock_zcopy(struct socket *sock)
229 {
230 	return unlikely(experimental_zcopytx) &&
231 		sock_flag(sock->sk, SOCK_ZEROCOPY);
232 }
233 
234 /* In case of DMA done not in order in lower device driver for some reason.
235  * upend_idx is used to track end of used idx, done_idx is used to track head
236  * of used idx. Once lower device DMA done contiguously, we will signal KVM
237  * guest used idx.
238  */
239 static void vhost_zerocopy_signal_used(struct vhost_net *net,
240 				       struct vhost_virtqueue *vq)
241 {
242 	struct vhost_net_virtqueue *nvq =
243 		container_of(vq, struct vhost_net_virtqueue, vq);
244 	int i, add;
245 	int j = 0;
246 
247 	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
248 		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
249 			vhost_net_tx_err(net);
250 		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
251 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
252 			++j;
253 		} else
254 			break;
255 	}
256 	while (j) {
257 		add = min(UIO_MAXIOV - nvq->done_idx, j);
258 		vhost_add_used_and_signal_n(vq->dev, vq,
259 					    &vq->heads[nvq->done_idx], add);
260 		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
261 		j -= add;
262 	}
263 }
264 
265 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
266 {
267 	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
268 	struct vhost_virtqueue *vq = ubufs->vq;
269 	int cnt;
270 
271 	rcu_read_lock_bh();
272 
273 	/* set len to mark this desc buffers done DMA */
274 	vq->heads[ubuf->desc].len = success ?
275 		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
276 	cnt = vhost_net_ubuf_put(ubufs);
277 
278 	/*
279 	 * Trigger polling thread if guest stopped submitting new buffers:
280 	 * in this case, the refcount after decrement will eventually reach 1.
281 	 * We also trigger polling periodically after each 16 packets
282 	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
283 	 * less than 10% of times).
284 	 */
285 	if (cnt <= 1 || !(cnt % 16))
286 		vhost_poll_queue(&vq->poll);
287 
288 	rcu_read_unlock_bh();
289 }
290 
291 /* Expects to be always run from workqueue - which acts as
292  * read-size critical section for our kind of RCU. */
293 static void handle_tx(struct vhost_net *net)
294 {
295 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
296 	struct vhost_virtqueue *vq = &nvq->vq;
297 	unsigned out, in;
298 	int head;
299 	struct msghdr msg = {
300 		.msg_name = NULL,
301 		.msg_namelen = 0,
302 		.msg_control = NULL,
303 		.msg_controllen = 0,
304 		.msg_flags = MSG_DONTWAIT,
305 	};
306 	size_t len, total_len = 0;
307 	int err;
308 	size_t hdr_size;
309 	struct socket *sock;
310 	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
311 	bool zcopy, zcopy_used;
312 
313 	mutex_lock(&vq->mutex);
314 	sock = vq->private_data;
315 	if (!sock)
316 		goto out;
317 
318 	vhost_disable_notify(&net->dev, vq);
319 
320 	hdr_size = nvq->vhost_hlen;
321 	zcopy = nvq->ubufs;
322 
323 	for (;;) {
324 		/* Release DMAs done buffers first */
325 		if (zcopy)
326 			vhost_zerocopy_signal_used(net, vq);
327 
328 		/* If more outstanding DMAs, queue the work.
329 		 * Handle upend_idx wrap around
330 		 */
331 		if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
332 			      % UIO_MAXIOV == nvq->done_idx))
333 			break;
334 
335 		head = vhost_get_vq_desc(vq, vq->iov,
336 					 ARRAY_SIZE(vq->iov),
337 					 &out, &in,
338 					 NULL, NULL);
339 		/* On error, stop handling until the next kick. */
340 		if (unlikely(head < 0))
341 			break;
342 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
343 		if (head == vq->num) {
344 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
345 				vhost_disable_notify(&net->dev, vq);
346 				continue;
347 			}
348 			break;
349 		}
350 		if (in) {
351 			vq_err(vq, "Unexpected descriptor format for TX: "
352 			       "out %d, int %d\n", out, in);
353 			break;
354 		}
355 		/* Skip header. TODO: support TSO. */
356 		len = iov_length(vq->iov, out);
357 		iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
358 		iov_iter_advance(&msg.msg_iter, hdr_size);
359 		/* Sanity check */
360 		if (!iov_iter_count(&msg.msg_iter)) {
361 			vq_err(vq, "Unexpected header len for TX: "
362 			       "%zd expected %zd\n",
363 			       len, hdr_size);
364 			break;
365 		}
366 		len = iov_iter_count(&msg.msg_iter);
367 
368 		zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
369 				   && (nvq->upend_idx + 1) % UIO_MAXIOV !=
370 				      nvq->done_idx
371 				   && vhost_net_tx_select_zcopy(net);
372 
373 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
374 		if (zcopy_used) {
375 			struct ubuf_info *ubuf;
376 			ubuf = nvq->ubuf_info + nvq->upend_idx;
377 
378 			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
379 			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
380 			ubuf->callback = vhost_zerocopy_callback;
381 			ubuf->ctx = nvq->ubufs;
382 			ubuf->desc = nvq->upend_idx;
383 			msg.msg_control = ubuf;
384 			msg.msg_controllen = sizeof(ubuf);
385 			ubufs = nvq->ubufs;
386 			atomic_inc(&ubufs->refcount);
387 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
388 		} else {
389 			msg.msg_control = NULL;
390 			ubufs = NULL;
391 		}
392 		/* TODO: Check specific error and bomb out unless ENOBUFS? */
393 		err = sock->ops->sendmsg(NULL, sock, &msg, len);
394 		if (unlikely(err < 0)) {
395 			if (zcopy_used) {
396 				vhost_net_ubuf_put(ubufs);
397 				nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
398 					% UIO_MAXIOV;
399 			}
400 			vhost_discard_vq_desc(vq, 1);
401 			break;
402 		}
403 		if (err != len)
404 			pr_debug("Truncated TX packet: "
405 				 " len %d != %zd\n", err, len);
406 		if (!zcopy_used)
407 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
408 		else
409 			vhost_zerocopy_signal_used(net, vq);
410 		total_len += len;
411 		vhost_net_tx_packet(net);
412 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
413 			vhost_poll_queue(&vq->poll);
414 			break;
415 		}
416 	}
417 out:
418 	mutex_unlock(&vq->mutex);
419 }
420 
421 static int peek_head_len(struct sock *sk)
422 {
423 	struct sk_buff *head;
424 	int len = 0;
425 	unsigned long flags;
426 
427 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
428 	head = skb_peek(&sk->sk_receive_queue);
429 	if (likely(head)) {
430 		len = head->len;
431 		if (skb_vlan_tag_present(head))
432 			len += VLAN_HLEN;
433 	}
434 
435 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
436 	return len;
437 }
438 
439 /* This is a multi-buffer version of vhost_get_desc, that works if
440  *	vq has read descriptors only.
441  * @vq		- the relevant virtqueue
442  * @datalen	- data length we'll be reading
443  * @iovcount	- returned count of io vectors we fill
444  * @log		- vhost log
445  * @log_num	- log offset
446  * @quota       - headcount quota, 1 for big buffer
447  *	returns number of buffer heads allocated, negative on error
448  */
449 static int get_rx_bufs(struct vhost_virtqueue *vq,
450 		       struct vring_used_elem *heads,
451 		       int datalen,
452 		       unsigned *iovcount,
453 		       struct vhost_log *log,
454 		       unsigned *log_num,
455 		       unsigned int quota)
456 {
457 	unsigned int out, in;
458 	int seg = 0;
459 	int headcount = 0;
460 	unsigned d;
461 	int r, nlogs = 0;
462 	/* len is always initialized before use since we are always called with
463 	 * datalen > 0.
464 	 */
465 	u32 uninitialized_var(len);
466 
467 	while (datalen > 0 && headcount < quota) {
468 		if (unlikely(seg >= UIO_MAXIOV)) {
469 			r = -ENOBUFS;
470 			goto err;
471 		}
472 		r = vhost_get_vq_desc(vq, vq->iov + seg,
473 				      ARRAY_SIZE(vq->iov) - seg, &out,
474 				      &in, log, log_num);
475 		if (unlikely(r < 0))
476 			goto err;
477 
478 		d = r;
479 		if (d == vq->num) {
480 			r = 0;
481 			goto err;
482 		}
483 		if (unlikely(out || in <= 0)) {
484 			vq_err(vq, "unexpected descriptor format for RX: "
485 				"out %d, in %d\n", out, in);
486 			r = -EINVAL;
487 			goto err;
488 		}
489 		if (unlikely(log)) {
490 			nlogs += *log_num;
491 			log += *log_num;
492 		}
493 		heads[headcount].id = cpu_to_vhost32(vq, d);
494 		len = iov_length(vq->iov + seg, in);
495 		heads[headcount].len = cpu_to_vhost32(vq, len);
496 		datalen -= len;
497 		++headcount;
498 		seg += in;
499 	}
500 	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
501 	*iovcount = seg;
502 	if (unlikely(log))
503 		*log_num = nlogs;
504 
505 	/* Detect overrun */
506 	if (unlikely(datalen > 0)) {
507 		r = UIO_MAXIOV + 1;
508 		goto err;
509 	}
510 	return headcount;
511 err:
512 	vhost_discard_vq_desc(vq, headcount);
513 	return r;
514 }
515 
516 /* Expects to be always run from workqueue - which acts as
517  * read-size critical section for our kind of RCU. */
518 static void handle_rx(struct vhost_net *net)
519 {
520 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
521 	struct vhost_virtqueue *vq = &nvq->vq;
522 	unsigned uninitialized_var(in), log;
523 	struct vhost_log *vq_log;
524 	struct msghdr msg = {
525 		.msg_name = NULL,
526 		.msg_namelen = 0,
527 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
528 		.msg_controllen = 0,
529 		.msg_flags = MSG_DONTWAIT,
530 	};
531 	struct virtio_net_hdr hdr = {
532 		.flags = 0,
533 		.gso_type = VIRTIO_NET_HDR_GSO_NONE
534 	};
535 	size_t total_len = 0;
536 	int err, mergeable;
537 	s16 headcount;
538 	size_t vhost_hlen, sock_hlen;
539 	size_t vhost_len, sock_len;
540 	struct socket *sock;
541 	struct iov_iter fixup;
542 	__virtio16 num_buffers;
543 
544 	mutex_lock(&vq->mutex);
545 	sock = vq->private_data;
546 	if (!sock)
547 		goto out;
548 	vhost_disable_notify(&net->dev, vq);
549 
550 	vhost_hlen = nvq->vhost_hlen;
551 	sock_hlen = nvq->sock_hlen;
552 
553 	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
554 		vq->log : NULL;
555 	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
556 
557 	while ((sock_len = peek_head_len(sock->sk))) {
558 		sock_len += sock_hlen;
559 		vhost_len = sock_len + vhost_hlen;
560 		headcount = get_rx_bufs(vq, vq->heads, vhost_len,
561 					&in, vq_log, &log,
562 					likely(mergeable) ? UIO_MAXIOV : 1);
563 		/* On error, stop handling until the next kick. */
564 		if (unlikely(headcount < 0))
565 			break;
566 		/* On overrun, truncate and discard */
567 		if (unlikely(headcount > UIO_MAXIOV)) {
568 			iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
569 			err = sock->ops->recvmsg(NULL, sock, &msg,
570 						 1, MSG_DONTWAIT | MSG_TRUNC);
571 			pr_debug("Discarded rx packet: len %zd\n", sock_len);
572 			continue;
573 		}
574 		/* OK, now we need to know about added descriptors. */
575 		if (!headcount) {
576 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
577 				/* They have slipped one in as we were
578 				 * doing that: check again. */
579 				vhost_disable_notify(&net->dev, vq);
580 				continue;
581 			}
582 			/* Nothing new?  Wait for eventfd to tell us
583 			 * they refilled. */
584 			break;
585 		}
586 		/* We don't need to be notified again. */
587 		iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
588 		fixup = msg.msg_iter;
589 		if (unlikely((vhost_hlen))) {
590 			/* We will supply the header ourselves
591 			 * TODO: support TSO.
592 			 */
593 			iov_iter_advance(&msg.msg_iter, vhost_hlen);
594 		}
595 		err = sock->ops->recvmsg(NULL, sock, &msg,
596 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
597 		/* Userspace might have consumed the packet meanwhile:
598 		 * it's not supposed to do this usually, but might be hard
599 		 * to prevent. Discard data we got (if any) and keep going. */
600 		if (unlikely(err != sock_len)) {
601 			pr_debug("Discarded rx packet: "
602 				 " len %d, expected %zd\n", err, sock_len);
603 			vhost_discard_vq_desc(vq, headcount);
604 			continue;
605 		}
606 		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
607 		if (unlikely(vhost_hlen)) {
608 			if (copy_to_iter(&hdr, sizeof(hdr),
609 					 &fixup) != sizeof(hdr)) {
610 				vq_err(vq, "Unable to write vnet_hdr "
611 				       "at addr %p\n", vq->iov->iov_base);
612 				break;
613 			}
614 		} else {
615 			/* Header came from socket; we'll need to patch
616 			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
617 			 */
618 			iov_iter_advance(&fixup, sizeof(hdr));
619 		}
620 		/* TODO: Should check and handle checksum. */
621 
622 		num_buffers = cpu_to_vhost16(vq, headcount);
623 		if (likely(mergeable) &&
624 		    copy_to_iter(&num_buffers, sizeof num_buffers,
625 				 &fixup) != sizeof num_buffers) {
626 			vq_err(vq, "Failed num_buffers write");
627 			vhost_discard_vq_desc(vq, headcount);
628 			break;
629 		}
630 		vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
631 					    headcount);
632 		if (unlikely(vq_log))
633 			vhost_log_write(vq, vq_log, log, vhost_len);
634 		total_len += vhost_len;
635 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
636 			vhost_poll_queue(&vq->poll);
637 			break;
638 		}
639 	}
640 out:
641 	mutex_unlock(&vq->mutex);
642 }
643 
644 static void handle_tx_kick(struct vhost_work *work)
645 {
646 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
647 						  poll.work);
648 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
649 
650 	handle_tx(net);
651 }
652 
653 static void handle_rx_kick(struct vhost_work *work)
654 {
655 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
656 						  poll.work);
657 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
658 
659 	handle_rx(net);
660 }
661 
662 static void handle_tx_net(struct vhost_work *work)
663 {
664 	struct vhost_net *net = container_of(work, struct vhost_net,
665 					     poll[VHOST_NET_VQ_TX].work);
666 	handle_tx(net);
667 }
668 
669 static void handle_rx_net(struct vhost_work *work)
670 {
671 	struct vhost_net *net = container_of(work, struct vhost_net,
672 					     poll[VHOST_NET_VQ_RX].work);
673 	handle_rx(net);
674 }
675 
676 static int vhost_net_open(struct inode *inode, struct file *f)
677 {
678 	struct vhost_net *n;
679 	struct vhost_dev *dev;
680 	struct vhost_virtqueue **vqs;
681 	int i;
682 
683 	n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
684 	if (!n) {
685 		n = vmalloc(sizeof *n);
686 		if (!n)
687 			return -ENOMEM;
688 	}
689 	vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
690 	if (!vqs) {
691 		kvfree(n);
692 		return -ENOMEM;
693 	}
694 
695 	dev = &n->dev;
696 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
697 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
698 	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
699 	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
700 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
701 		n->vqs[i].ubufs = NULL;
702 		n->vqs[i].ubuf_info = NULL;
703 		n->vqs[i].upend_idx = 0;
704 		n->vqs[i].done_idx = 0;
705 		n->vqs[i].vhost_hlen = 0;
706 		n->vqs[i].sock_hlen = 0;
707 	}
708 	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
709 
710 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
711 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
712 
713 	f->private_data = n;
714 
715 	return 0;
716 }
717 
718 static void vhost_net_disable_vq(struct vhost_net *n,
719 				 struct vhost_virtqueue *vq)
720 {
721 	struct vhost_net_virtqueue *nvq =
722 		container_of(vq, struct vhost_net_virtqueue, vq);
723 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
724 	if (!vq->private_data)
725 		return;
726 	vhost_poll_stop(poll);
727 }
728 
729 static int vhost_net_enable_vq(struct vhost_net *n,
730 				struct vhost_virtqueue *vq)
731 {
732 	struct vhost_net_virtqueue *nvq =
733 		container_of(vq, struct vhost_net_virtqueue, vq);
734 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
735 	struct socket *sock;
736 
737 	sock = vq->private_data;
738 	if (!sock)
739 		return 0;
740 
741 	return vhost_poll_start(poll, sock->file);
742 }
743 
744 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
745 					struct vhost_virtqueue *vq)
746 {
747 	struct socket *sock;
748 
749 	mutex_lock(&vq->mutex);
750 	sock = vq->private_data;
751 	vhost_net_disable_vq(n, vq);
752 	vq->private_data = NULL;
753 	mutex_unlock(&vq->mutex);
754 	return sock;
755 }
756 
757 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
758 			   struct socket **rx_sock)
759 {
760 	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
761 	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
762 }
763 
764 static void vhost_net_flush_vq(struct vhost_net *n, int index)
765 {
766 	vhost_poll_flush(n->poll + index);
767 	vhost_poll_flush(&n->vqs[index].vq.poll);
768 }
769 
770 static void vhost_net_flush(struct vhost_net *n)
771 {
772 	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
773 	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
774 	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
775 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
776 		n->tx_flush = true;
777 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
778 		/* Wait for all lower device DMAs done. */
779 		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
780 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
781 		n->tx_flush = false;
782 		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
783 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
784 	}
785 }
786 
787 static int vhost_net_release(struct inode *inode, struct file *f)
788 {
789 	struct vhost_net *n = f->private_data;
790 	struct socket *tx_sock;
791 	struct socket *rx_sock;
792 
793 	vhost_net_stop(n, &tx_sock, &rx_sock);
794 	vhost_net_flush(n);
795 	vhost_dev_stop(&n->dev);
796 	vhost_dev_cleanup(&n->dev, false);
797 	vhost_net_vq_reset(n);
798 	if (tx_sock)
799 		sockfd_put(tx_sock);
800 	if (rx_sock)
801 		sockfd_put(rx_sock);
802 	/* Make sure no callbacks are outstanding */
803 	synchronize_rcu_bh();
804 	/* We do an extra flush before freeing memory,
805 	 * since jobs can re-queue themselves. */
806 	vhost_net_flush(n);
807 	kfree(n->dev.vqs);
808 	kvfree(n);
809 	return 0;
810 }
811 
812 static struct socket *get_raw_socket(int fd)
813 {
814 	struct {
815 		struct sockaddr_ll sa;
816 		char  buf[MAX_ADDR_LEN];
817 	} uaddr;
818 	int uaddr_len = sizeof uaddr, r;
819 	struct socket *sock = sockfd_lookup(fd, &r);
820 
821 	if (!sock)
822 		return ERR_PTR(-ENOTSOCK);
823 
824 	/* Parameter checking */
825 	if (sock->sk->sk_type != SOCK_RAW) {
826 		r = -ESOCKTNOSUPPORT;
827 		goto err;
828 	}
829 
830 	r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
831 			       &uaddr_len, 0);
832 	if (r)
833 		goto err;
834 
835 	if (uaddr.sa.sll_family != AF_PACKET) {
836 		r = -EPFNOSUPPORT;
837 		goto err;
838 	}
839 	return sock;
840 err:
841 	sockfd_put(sock);
842 	return ERR_PTR(r);
843 }
844 
845 static struct socket *get_tap_socket(int fd)
846 {
847 	struct file *file = fget(fd);
848 	struct socket *sock;
849 
850 	if (!file)
851 		return ERR_PTR(-EBADF);
852 	sock = tun_get_socket(file);
853 	if (!IS_ERR(sock))
854 		return sock;
855 	sock = macvtap_get_socket(file);
856 	if (IS_ERR(sock))
857 		fput(file);
858 	return sock;
859 }
860 
861 static struct socket *get_socket(int fd)
862 {
863 	struct socket *sock;
864 
865 	/* special case to disable backend */
866 	if (fd == -1)
867 		return NULL;
868 	sock = get_raw_socket(fd);
869 	if (!IS_ERR(sock))
870 		return sock;
871 	sock = get_tap_socket(fd);
872 	if (!IS_ERR(sock))
873 		return sock;
874 	return ERR_PTR(-ENOTSOCK);
875 }
876 
877 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
878 {
879 	struct socket *sock, *oldsock;
880 	struct vhost_virtqueue *vq;
881 	struct vhost_net_virtqueue *nvq;
882 	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
883 	int r;
884 
885 	mutex_lock(&n->dev.mutex);
886 	r = vhost_dev_check_owner(&n->dev);
887 	if (r)
888 		goto err;
889 
890 	if (index >= VHOST_NET_VQ_MAX) {
891 		r = -ENOBUFS;
892 		goto err;
893 	}
894 	vq = &n->vqs[index].vq;
895 	nvq = &n->vqs[index];
896 	mutex_lock(&vq->mutex);
897 
898 	/* Verify that ring has been setup correctly. */
899 	if (!vhost_vq_access_ok(vq)) {
900 		r = -EFAULT;
901 		goto err_vq;
902 	}
903 	sock = get_socket(fd);
904 	if (IS_ERR(sock)) {
905 		r = PTR_ERR(sock);
906 		goto err_vq;
907 	}
908 
909 	/* start polling new socket */
910 	oldsock = vq->private_data;
911 	if (sock != oldsock) {
912 		ubufs = vhost_net_ubuf_alloc(vq,
913 					     sock && vhost_sock_zcopy(sock));
914 		if (IS_ERR(ubufs)) {
915 			r = PTR_ERR(ubufs);
916 			goto err_ubufs;
917 		}
918 
919 		vhost_net_disable_vq(n, vq);
920 		vq->private_data = sock;
921 		r = vhost_init_used(vq);
922 		if (r)
923 			goto err_used;
924 		r = vhost_net_enable_vq(n, vq);
925 		if (r)
926 			goto err_used;
927 
928 		oldubufs = nvq->ubufs;
929 		nvq->ubufs = ubufs;
930 
931 		n->tx_packets = 0;
932 		n->tx_zcopy_err = 0;
933 		n->tx_flush = false;
934 	}
935 
936 	mutex_unlock(&vq->mutex);
937 
938 	if (oldubufs) {
939 		vhost_net_ubuf_put_wait_and_free(oldubufs);
940 		mutex_lock(&vq->mutex);
941 		vhost_zerocopy_signal_used(n, vq);
942 		mutex_unlock(&vq->mutex);
943 	}
944 
945 	if (oldsock) {
946 		vhost_net_flush_vq(n, index);
947 		sockfd_put(oldsock);
948 	}
949 
950 	mutex_unlock(&n->dev.mutex);
951 	return 0;
952 
953 err_used:
954 	vq->private_data = oldsock;
955 	vhost_net_enable_vq(n, vq);
956 	if (ubufs)
957 		vhost_net_ubuf_put_wait_and_free(ubufs);
958 err_ubufs:
959 	sockfd_put(sock);
960 err_vq:
961 	mutex_unlock(&vq->mutex);
962 err:
963 	mutex_unlock(&n->dev.mutex);
964 	return r;
965 }
966 
967 static long vhost_net_reset_owner(struct vhost_net *n)
968 {
969 	struct socket *tx_sock = NULL;
970 	struct socket *rx_sock = NULL;
971 	long err;
972 	struct vhost_memory *memory;
973 
974 	mutex_lock(&n->dev.mutex);
975 	err = vhost_dev_check_owner(&n->dev);
976 	if (err)
977 		goto done;
978 	memory = vhost_dev_reset_owner_prepare();
979 	if (!memory) {
980 		err = -ENOMEM;
981 		goto done;
982 	}
983 	vhost_net_stop(n, &tx_sock, &rx_sock);
984 	vhost_net_flush(n);
985 	vhost_dev_reset_owner(&n->dev, memory);
986 	vhost_net_vq_reset(n);
987 done:
988 	mutex_unlock(&n->dev.mutex);
989 	if (tx_sock)
990 		sockfd_put(tx_sock);
991 	if (rx_sock)
992 		sockfd_put(rx_sock);
993 	return err;
994 }
995 
996 static int vhost_net_set_features(struct vhost_net *n, u64 features)
997 {
998 	size_t vhost_hlen, sock_hlen, hdr_len;
999 	int i;
1000 
1001 	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1002 			       (1ULL << VIRTIO_F_VERSION_1))) ?
1003 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1004 			sizeof(struct virtio_net_hdr);
1005 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1006 		/* vhost provides vnet_hdr */
1007 		vhost_hlen = hdr_len;
1008 		sock_hlen = 0;
1009 	} else {
1010 		/* socket provides vnet_hdr */
1011 		vhost_hlen = 0;
1012 		sock_hlen = hdr_len;
1013 	}
1014 	mutex_lock(&n->dev.mutex);
1015 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1016 	    !vhost_log_access_ok(&n->dev)) {
1017 		mutex_unlock(&n->dev.mutex);
1018 		return -EFAULT;
1019 	}
1020 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1021 		mutex_lock(&n->vqs[i].vq.mutex);
1022 		n->vqs[i].vq.acked_features = features;
1023 		n->vqs[i].vhost_hlen = vhost_hlen;
1024 		n->vqs[i].sock_hlen = sock_hlen;
1025 		mutex_unlock(&n->vqs[i].vq.mutex);
1026 	}
1027 	mutex_unlock(&n->dev.mutex);
1028 	return 0;
1029 }
1030 
1031 static long vhost_net_set_owner(struct vhost_net *n)
1032 {
1033 	int r;
1034 
1035 	mutex_lock(&n->dev.mutex);
1036 	if (vhost_dev_has_owner(&n->dev)) {
1037 		r = -EBUSY;
1038 		goto out;
1039 	}
1040 	r = vhost_net_set_ubuf_info(n);
1041 	if (r)
1042 		goto out;
1043 	r = vhost_dev_set_owner(&n->dev);
1044 	if (r)
1045 		vhost_net_clear_ubuf_info(n);
1046 	vhost_net_flush(n);
1047 out:
1048 	mutex_unlock(&n->dev.mutex);
1049 	return r;
1050 }
1051 
1052 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1053 			    unsigned long arg)
1054 {
1055 	struct vhost_net *n = f->private_data;
1056 	void __user *argp = (void __user *)arg;
1057 	u64 __user *featurep = argp;
1058 	struct vhost_vring_file backend;
1059 	u64 features;
1060 	int r;
1061 
1062 	switch (ioctl) {
1063 	case VHOST_NET_SET_BACKEND:
1064 		if (copy_from_user(&backend, argp, sizeof backend))
1065 			return -EFAULT;
1066 		return vhost_net_set_backend(n, backend.index, backend.fd);
1067 	case VHOST_GET_FEATURES:
1068 		features = VHOST_NET_FEATURES;
1069 		if (copy_to_user(featurep, &features, sizeof features))
1070 			return -EFAULT;
1071 		return 0;
1072 	case VHOST_SET_FEATURES:
1073 		if (copy_from_user(&features, featurep, sizeof features))
1074 			return -EFAULT;
1075 		if (features & ~VHOST_NET_FEATURES)
1076 			return -EOPNOTSUPP;
1077 		return vhost_net_set_features(n, features);
1078 	case VHOST_RESET_OWNER:
1079 		return vhost_net_reset_owner(n);
1080 	case VHOST_SET_OWNER:
1081 		return vhost_net_set_owner(n);
1082 	default:
1083 		mutex_lock(&n->dev.mutex);
1084 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1085 		if (r == -ENOIOCTLCMD)
1086 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1087 		else
1088 			vhost_net_flush(n);
1089 		mutex_unlock(&n->dev.mutex);
1090 		return r;
1091 	}
1092 }
1093 
1094 #ifdef CONFIG_COMPAT
1095 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1096 				   unsigned long arg)
1097 {
1098 	return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1099 }
1100 #endif
1101 
1102 static const struct file_operations vhost_net_fops = {
1103 	.owner          = THIS_MODULE,
1104 	.release        = vhost_net_release,
1105 	.unlocked_ioctl = vhost_net_ioctl,
1106 #ifdef CONFIG_COMPAT
1107 	.compat_ioctl   = vhost_net_compat_ioctl,
1108 #endif
1109 	.open           = vhost_net_open,
1110 	.llseek		= noop_llseek,
1111 };
1112 
1113 static struct miscdevice vhost_net_misc = {
1114 	.minor = VHOST_NET_MINOR,
1115 	.name = "vhost-net",
1116 	.fops = &vhost_net_fops,
1117 };
1118 
1119 static int vhost_net_init(void)
1120 {
1121 	if (experimental_zcopytx)
1122 		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1123 	return misc_register(&vhost_net_misc);
1124 }
1125 module_init(vhost_net_init);
1126 
1127 static void vhost_net_exit(void)
1128 {
1129 	misc_deregister(&vhost_net_misc);
1130 }
1131 module_exit(vhost_net_exit);
1132 
1133 MODULE_VERSION("0.0.1");
1134 MODULE_LICENSE("GPL v2");
1135 MODULE_AUTHOR("Michael S. Tsirkin");
1136 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1137 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1138 MODULE_ALIAS("devname:vhost-net");
1139