xref: /linux/drivers/vhost/net.c (revision 54f5a57e266318d72f84fda95805099986a7e201)
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 
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 #include <linux/if_vlan.h>
27 
28 #include <net/sock.h>
29 
30 #include "vhost.h"
31 
32 static int experimental_zcopytx = 1;
33 module_param(experimental_zcopytx, int, 0444);
34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
35 		                       " 1 -Enable; 0 - Disable");
36 
37 /* Max number of bytes transferred before requeueing the job.
38  * Using this limit prevents one virtqueue from starving others. */
39 #define VHOST_NET_WEIGHT 0x80000
40 
41 /* MAX number of TX used buffers for outstanding zerocopy */
42 #define VHOST_MAX_PEND 128
43 #define VHOST_GOODCOPY_LEN 256
44 
45 /*
46  * For transmit, used buffer len is unused; we override it to track buffer
47  * status internally; used for zerocopy tx only.
48  */
49 /* Lower device DMA failed */
50 #define VHOST_DMA_FAILED_LEN	3
51 /* Lower device DMA done */
52 #define VHOST_DMA_DONE_LEN	2
53 /* Lower device DMA in progress */
54 #define VHOST_DMA_IN_PROGRESS	1
55 /* Buffer unused */
56 #define VHOST_DMA_CLEAR_LEN	0
57 
58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
59 
60 enum {
61 	VHOST_NET_FEATURES = VHOST_FEATURES |
62 			 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
63 			 (1ULL << VIRTIO_NET_F_MRG_RXBUF),
64 };
65 
66 enum {
67 	VHOST_NET_VQ_RX = 0,
68 	VHOST_NET_VQ_TX = 1,
69 	VHOST_NET_VQ_MAX = 2,
70 };
71 
72 struct vhost_net_ubuf_ref {
73 	/* refcount follows semantics similar to kref:
74 	 *  0: object is released
75 	 *  1: no outstanding ubufs
76 	 * >1: outstanding ubufs
77 	 */
78 	atomic_t refcount;
79 	wait_queue_head_t wait;
80 	struct vhost_virtqueue *vq;
81 };
82 
83 struct vhost_net_virtqueue {
84 	struct vhost_virtqueue vq;
85 	/* hdr is used to store the virtio header.
86 	 * Since each iovec has >= 1 byte length, we never need more than
87 	 * header length entries to store the header. */
88 	struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
89 	size_t vhost_hlen;
90 	size_t sock_hlen;
91 	/* vhost zerocopy support fields below: */
92 	/* last used idx for outstanding DMA zerocopy buffers */
93 	int upend_idx;
94 	/* first used idx for DMA done zerocopy buffers */
95 	int done_idx;
96 	/* an array of userspace buffers info */
97 	struct ubuf_info *ubuf_info;
98 	/* Reference counting for outstanding ubufs.
99 	 * Protected by vq mutex. Writers must also take device mutex. */
100 	struct vhost_net_ubuf_ref *ubufs;
101 };
102 
103 struct vhost_net {
104 	struct vhost_dev dev;
105 	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
106 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
107 	/* Number of TX recently submitted.
108 	 * Protected by tx vq lock. */
109 	unsigned tx_packets;
110 	/* Number of times zerocopy TX recently failed.
111 	 * Protected by tx vq lock. */
112 	unsigned tx_zcopy_err;
113 	/* Flush in progress. Protected by tx vq lock. */
114 	bool tx_flush;
115 };
116 
117 static unsigned vhost_net_zcopy_mask __read_mostly;
118 
119 static void vhost_net_enable_zcopy(int vq)
120 {
121 	vhost_net_zcopy_mask |= 0x1 << vq;
122 }
123 
124 static struct vhost_net_ubuf_ref *
125 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
126 {
127 	struct vhost_net_ubuf_ref *ubufs;
128 	/* No zero copy backend? Nothing to count. */
129 	if (!zcopy)
130 		return NULL;
131 	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
132 	if (!ubufs)
133 		return ERR_PTR(-ENOMEM);
134 	atomic_set(&ubufs->refcount, 1);
135 	init_waitqueue_head(&ubufs->wait);
136 	ubufs->vq = vq;
137 	return ubufs;
138 }
139 
140 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
141 {
142 	int r = atomic_sub_return(1, &ubufs->refcount);
143 	if (unlikely(!r))
144 		wake_up(&ubufs->wait);
145 	return r;
146 }
147 
148 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
149 {
150 	vhost_net_ubuf_put(ubufs);
151 	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
152 }
153 
154 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
155 {
156 	vhost_net_ubuf_put_and_wait(ubufs);
157 	kfree(ubufs);
158 }
159 
160 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
161 {
162 	int i;
163 
164 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
165 		kfree(n->vqs[i].ubuf_info);
166 		n->vqs[i].ubuf_info = NULL;
167 	}
168 }
169 
170 static int vhost_net_set_ubuf_info(struct vhost_net *n)
171 {
172 	bool zcopy;
173 	int i;
174 
175 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
176 		zcopy = vhost_net_zcopy_mask & (0x1 << i);
177 		if (!zcopy)
178 			continue;
179 		n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
180 					      UIO_MAXIOV, GFP_KERNEL);
181 		if  (!n->vqs[i].ubuf_info)
182 			goto err;
183 	}
184 	return 0;
185 
186 err:
187 	vhost_net_clear_ubuf_info(n);
188 	return -ENOMEM;
189 }
190 
191 static void vhost_net_vq_reset(struct vhost_net *n)
192 {
193 	int i;
194 
195 	vhost_net_clear_ubuf_info(n);
196 
197 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
198 		n->vqs[i].done_idx = 0;
199 		n->vqs[i].upend_idx = 0;
200 		n->vqs[i].ubufs = NULL;
201 		n->vqs[i].vhost_hlen = 0;
202 		n->vqs[i].sock_hlen = 0;
203 	}
204 
205 }
206 
207 static void vhost_net_tx_packet(struct vhost_net *net)
208 {
209 	++net->tx_packets;
210 	if (net->tx_packets < 1024)
211 		return;
212 	net->tx_packets = 0;
213 	net->tx_zcopy_err = 0;
214 }
215 
216 static void vhost_net_tx_err(struct vhost_net *net)
217 {
218 	++net->tx_zcopy_err;
219 }
220 
221 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
222 {
223 	/* TX flush waits for outstanding DMAs to be done.
224 	 * Don't start new DMAs.
225 	 */
226 	return !net->tx_flush &&
227 		net->tx_packets / 64 >= net->tx_zcopy_err;
228 }
229 
230 static bool vhost_sock_zcopy(struct socket *sock)
231 {
232 	return unlikely(experimental_zcopytx) &&
233 		sock_flag(sock->sk, SOCK_ZEROCOPY);
234 }
235 
236 /* Pop first len bytes from iovec. Return number of segments used. */
237 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
238 			  size_t len, int iov_count)
239 {
240 	int seg = 0;
241 	size_t size;
242 
243 	while (len && seg < iov_count) {
244 		size = min(from->iov_len, len);
245 		to->iov_base = from->iov_base;
246 		to->iov_len = size;
247 		from->iov_len -= size;
248 		from->iov_base += size;
249 		len -= size;
250 		++from;
251 		++to;
252 		++seg;
253 	}
254 	return seg;
255 }
256 /* Copy iovec entries for len bytes from iovec. */
257 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
258 			   size_t len, int iovcount)
259 {
260 	int seg = 0;
261 	size_t size;
262 
263 	while (len && seg < iovcount) {
264 		size = min(from->iov_len, len);
265 		to->iov_base = from->iov_base;
266 		to->iov_len = size;
267 		len -= size;
268 		++from;
269 		++to;
270 		++seg;
271 	}
272 }
273 
274 /* In case of DMA done not in order in lower device driver for some reason.
275  * upend_idx is used to track end of used idx, done_idx is used to track head
276  * of used idx. Once lower device DMA done contiguously, we will signal KVM
277  * guest used idx.
278  */
279 static void vhost_zerocopy_signal_used(struct vhost_net *net,
280 				       struct vhost_virtqueue *vq)
281 {
282 	struct vhost_net_virtqueue *nvq =
283 		container_of(vq, struct vhost_net_virtqueue, vq);
284 	int i, add;
285 	int j = 0;
286 
287 	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
288 		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
289 			vhost_net_tx_err(net);
290 		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
291 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
292 			++j;
293 		} else
294 			break;
295 	}
296 	while (j) {
297 		add = min(UIO_MAXIOV - nvq->done_idx, j);
298 		vhost_add_used_and_signal_n(vq->dev, vq,
299 					    &vq->heads[nvq->done_idx], add);
300 		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
301 		j -= add;
302 	}
303 }
304 
305 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
306 {
307 	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
308 	struct vhost_virtqueue *vq = ubufs->vq;
309 	int cnt;
310 
311 	rcu_read_lock_bh();
312 
313 	/* set len to mark this desc buffers done DMA */
314 	vq->heads[ubuf->desc].len = success ?
315 		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
316 	cnt = vhost_net_ubuf_put(ubufs);
317 
318 	/*
319 	 * Trigger polling thread if guest stopped submitting new buffers:
320 	 * in this case, the refcount after decrement will eventually reach 1.
321 	 * We also trigger polling periodically after each 16 packets
322 	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
323 	 * less than 10% of times).
324 	 */
325 	if (cnt <= 1 || !(cnt % 16))
326 		vhost_poll_queue(&vq->poll);
327 
328 	rcu_read_unlock_bh();
329 }
330 
331 /* Expects to be always run from workqueue - which acts as
332  * read-size critical section for our kind of RCU. */
333 static void handle_tx(struct vhost_net *net)
334 {
335 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
336 	struct vhost_virtqueue *vq = &nvq->vq;
337 	unsigned out, in, s;
338 	int head;
339 	struct msghdr msg = {
340 		.msg_name = NULL,
341 		.msg_namelen = 0,
342 		.msg_control = NULL,
343 		.msg_controllen = 0,
344 		.msg_iov = vq->iov,
345 		.msg_flags = MSG_DONTWAIT,
346 	};
347 	size_t len, total_len = 0;
348 	int err;
349 	size_t hdr_size;
350 	struct socket *sock;
351 	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
352 	bool zcopy, zcopy_used;
353 
354 	mutex_lock(&vq->mutex);
355 	sock = vq->private_data;
356 	if (!sock)
357 		goto out;
358 
359 	vhost_disable_notify(&net->dev, vq);
360 
361 	hdr_size = nvq->vhost_hlen;
362 	zcopy = nvq->ubufs;
363 
364 	for (;;) {
365 		/* Release DMAs done buffers first */
366 		if (zcopy)
367 			vhost_zerocopy_signal_used(net, vq);
368 
369 		/* If more outstanding DMAs, queue the work.
370 		 * Handle upend_idx wrap around
371 		 */
372 		if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
373 			      % UIO_MAXIOV == nvq->done_idx))
374 			break;
375 
376 		head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
377 					 ARRAY_SIZE(vq->iov),
378 					 &out, &in,
379 					 NULL, NULL);
380 		/* On error, stop handling until the next kick. */
381 		if (unlikely(head < 0))
382 			break;
383 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
384 		if (head == vq->num) {
385 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
386 				vhost_disable_notify(&net->dev, vq);
387 				continue;
388 			}
389 			break;
390 		}
391 		if (in) {
392 			vq_err(vq, "Unexpected descriptor format for TX: "
393 			       "out %d, int %d\n", out, in);
394 			break;
395 		}
396 		/* Skip header. TODO: support TSO. */
397 		s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
398 		msg.msg_iovlen = out;
399 		len = iov_length(vq->iov, out);
400 		/* Sanity check */
401 		if (!len) {
402 			vq_err(vq, "Unexpected header len for TX: "
403 			       "%zd expected %zd\n",
404 			       iov_length(nvq->hdr, s), hdr_size);
405 			break;
406 		}
407 
408 		zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
409 				   && (nvq->upend_idx + 1) % UIO_MAXIOV !=
410 				      nvq->done_idx
411 				   && vhost_net_tx_select_zcopy(net);
412 
413 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
414 		if (zcopy_used) {
415 			struct ubuf_info *ubuf;
416 			ubuf = nvq->ubuf_info + nvq->upend_idx;
417 
418 			vq->heads[nvq->upend_idx].id = head;
419 			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
420 			ubuf->callback = vhost_zerocopy_callback;
421 			ubuf->ctx = nvq->ubufs;
422 			ubuf->desc = nvq->upend_idx;
423 			msg.msg_control = ubuf;
424 			msg.msg_controllen = sizeof(ubuf);
425 			ubufs = nvq->ubufs;
426 			atomic_inc(&ubufs->refcount);
427 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
428 		} else {
429 			msg.msg_control = NULL;
430 			ubufs = NULL;
431 		}
432 		/* TODO: Check specific error and bomb out unless ENOBUFS? */
433 		err = sock->ops->sendmsg(NULL, sock, &msg, len);
434 		if (unlikely(err < 0)) {
435 			if (zcopy_used) {
436 				vhost_net_ubuf_put(ubufs);
437 				nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
438 					% UIO_MAXIOV;
439 			}
440 			vhost_discard_vq_desc(vq, 1);
441 			break;
442 		}
443 		if (err != len)
444 			pr_debug("Truncated TX packet: "
445 				 " len %d != %zd\n", err, len);
446 		if (!zcopy_used)
447 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
448 		else
449 			vhost_zerocopy_signal_used(net, vq);
450 		total_len += len;
451 		vhost_net_tx_packet(net);
452 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
453 			vhost_poll_queue(&vq->poll);
454 			break;
455 		}
456 	}
457 out:
458 	mutex_unlock(&vq->mutex);
459 }
460 
461 static int peek_head_len(struct sock *sk)
462 {
463 	struct sk_buff *head;
464 	int len = 0;
465 	unsigned long flags;
466 
467 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
468 	head = skb_peek(&sk->sk_receive_queue);
469 	if (likely(head)) {
470 		len = head->len;
471 		if (vlan_tx_tag_present(head))
472 			len += VLAN_HLEN;
473 	}
474 
475 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
476 	return len;
477 }
478 
479 /* This is a multi-buffer version of vhost_get_desc, that works if
480  *	vq has read descriptors only.
481  * @vq		- the relevant virtqueue
482  * @datalen	- data length we'll be reading
483  * @iovcount	- returned count of io vectors we fill
484  * @log		- vhost log
485  * @log_num	- log offset
486  * @quota       - headcount quota, 1 for big buffer
487  *	returns number of buffer heads allocated, negative on error
488  */
489 static int get_rx_bufs(struct vhost_virtqueue *vq,
490 		       struct vring_used_elem *heads,
491 		       int datalen,
492 		       unsigned *iovcount,
493 		       struct vhost_log *log,
494 		       unsigned *log_num,
495 		       unsigned int quota)
496 {
497 	unsigned int out, in;
498 	int seg = 0;
499 	int headcount = 0;
500 	unsigned d;
501 	int r, nlogs = 0;
502 
503 	while (datalen > 0 && headcount < quota) {
504 		if (unlikely(seg >= UIO_MAXIOV)) {
505 			r = -ENOBUFS;
506 			goto err;
507 		}
508 		d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
509 				      ARRAY_SIZE(vq->iov) - seg, &out,
510 				      &in, log, log_num);
511 		if (d == vq->num) {
512 			r = 0;
513 			goto err;
514 		}
515 		if (unlikely(out || in <= 0)) {
516 			vq_err(vq, "unexpected descriptor format for RX: "
517 				"out %d, in %d\n", out, in);
518 			r = -EINVAL;
519 			goto err;
520 		}
521 		if (unlikely(log)) {
522 			nlogs += *log_num;
523 			log += *log_num;
524 		}
525 		heads[headcount].id = d;
526 		heads[headcount].len = iov_length(vq->iov + seg, in);
527 		datalen -= heads[headcount].len;
528 		++headcount;
529 		seg += in;
530 	}
531 	heads[headcount - 1].len += datalen;
532 	*iovcount = seg;
533 	if (unlikely(log))
534 		*log_num = nlogs;
535 	return headcount;
536 err:
537 	vhost_discard_vq_desc(vq, headcount);
538 	return r;
539 }
540 
541 /* Expects to be always run from workqueue - which acts as
542  * read-size critical section for our kind of RCU. */
543 static void handle_rx(struct vhost_net *net)
544 {
545 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
546 	struct vhost_virtqueue *vq = &nvq->vq;
547 	unsigned uninitialized_var(in), log;
548 	struct vhost_log *vq_log;
549 	struct msghdr msg = {
550 		.msg_name = NULL,
551 		.msg_namelen = 0,
552 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
553 		.msg_controllen = 0,
554 		.msg_iov = vq->iov,
555 		.msg_flags = MSG_DONTWAIT,
556 	};
557 	struct virtio_net_hdr_mrg_rxbuf hdr = {
558 		.hdr.flags = 0,
559 		.hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
560 	};
561 	size_t total_len = 0;
562 	int err, mergeable;
563 	s16 headcount;
564 	size_t vhost_hlen, sock_hlen;
565 	size_t vhost_len, sock_len;
566 	struct socket *sock;
567 
568 	mutex_lock(&vq->mutex);
569 	sock = vq->private_data;
570 	if (!sock)
571 		goto out;
572 	vhost_disable_notify(&net->dev, vq);
573 
574 	vhost_hlen = nvq->vhost_hlen;
575 	sock_hlen = nvq->sock_hlen;
576 
577 	vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
578 		vq->log : NULL;
579 	mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
580 
581 	while ((sock_len = peek_head_len(sock->sk))) {
582 		sock_len += sock_hlen;
583 		vhost_len = sock_len + vhost_hlen;
584 		headcount = get_rx_bufs(vq, vq->heads, vhost_len,
585 					&in, vq_log, &log,
586 					likely(mergeable) ? UIO_MAXIOV : 1);
587 		/* On error, stop handling until the next kick. */
588 		if (unlikely(headcount < 0))
589 			break;
590 		/* OK, now we need to know about added descriptors. */
591 		if (!headcount) {
592 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
593 				/* They have slipped one in as we were
594 				 * doing that: check again. */
595 				vhost_disable_notify(&net->dev, vq);
596 				continue;
597 			}
598 			/* Nothing new?  Wait for eventfd to tell us
599 			 * they refilled. */
600 			break;
601 		}
602 		/* We don't need to be notified again. */
603 		if (unlikely((vhost_hlen)))
604 			/* Skip header. TODO: support TSO. */
605 			move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
606 		else
607 			/* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
608 			 * needed because recvmsg can modify msg_iov. */
609 			copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
610 		msg.msg_iovlen = in;
611 		err = sock->ops->recvmsg(NULL, sock, &msg,
612 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
613 		/* Userspace might have consumed the packet meanwhile:
614 		 * it's not supposed to do this usually, but might be hard
615 		 * to prevent. Discard data we got (if any) and keep going. */
616 		if (unlikely(err != sock_len)) {
617 			pr_debug("Discarded rx packet: "
618 				 " len %d, expected %zd\n", err, sock_len);
619 			vhost_discard_vq_desc(vq, headcount);
620 			continue;
621 		}
622 		if (unlikely(vhost_hlen) &&
623 		    memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
624 				      vhost_hlen)) {
625 			vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
626 			       vq->iov->iov_base);
627 			break;
628 		}
629 		/* TODO: Should check and handle checksum. */
630 		if (likely(mergeable) &&
631 		    memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
632 				      offsetof(typeof(hdr), num_buffers),
633 				      sizeof hdr.num_buffers)) {
634 			vq_err(vq, "Failed num_buffers write");
635 			vhost_discard_vq_desc(vq, headcount);
636 			break;
637 		}
638 		vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
639 					    headcount);
640 		if (unlikely(vq_log))
641 			vhost_log_write(vq, vq_log, log, vhost_len);
642 		total_len += vhost_len;
643 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
644 			vhost_poll_queue(&vq->poll);
645 			break;
646 		}
647 	}
648 out:
649 	mutex_unlock(&vq->mutex);
650 }
651 
652 static void handle_tx_kick(struct vhost_work *work)
653 {
654 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
655 						  poll.work);
656 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
657 
658 	handle_tx(net);
659 }
660 
661 static void handle_rx_kick(struct vhost_work *work)
662 {
663 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
664 						  poll.work);
665 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
666 
667 	handle_rx(net);
668 }
669 
670 static void handle_tx_net(struct vhost_work *work)
671 {
672 	struct vhost_net *net = container_of(work, struct vhost_net,
673 					     poll[VHOST_NET_VQ_TX].work);
674 	handle_tx(net);
675 }
676 
677 static void handle_rx_net(struct vhost_work *work)
678 {
679 	struct vhost_net *net = container_of(work, struct vhost_net,
680 					     poll[VHOST_NET_VQ_RX].work);
681 	handle_rx(net);
682 }
683 
684 static int vhost_net_open(struct inode *inode, struct file *f)
685 {
686 	struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
687 	struct vhost_dev *dev;
688 	struct vhost_virtqueue **vqs;
689 	int i;
690 
691 	if (!n)
692 		return -ENOMEM;
693 	vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
694 	if (!vqs) {
695 		kfree(n);
696 		return -ENOMEM;
697 	}
698 
699 	dev = &n->dev;
700 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
701 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
702 	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
703 	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
704 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
705 		n->vqs[i].ubufs = NULL;
706 		n->vqs[i].ubuf_info = NULL;
707 		n->vqs[i].upend_idx = 0;
708 		n->vqs[i].done_idx = 0;
709 		n->vqs[i].vhost_hlen = 0;
710 		n->vqs[i].sock_hlen = 0;
711 	}
712 	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
713 
714 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
715 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
716 
717 	f->private_data = n;
718 
719 	return 0;
720 }
721 
722 static void vhost_net_disable_vq(struct vhost_net *n,
723 				 struct vhost_virtqueue *vq)
724 {
725 	struct vhost_net_virtqueue *nvq =
726 		container_of(vq, struct vhost_net_virtqueue, vq);
727 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
728 	if (!vq->private_data)
729 		return;
730 	vhost_poll_stop(poll);
731 }
732 
733 static int vhost_net_enable_vq(struct vhost_net *n,
734 				struct vhost_virtqueue *vq)
735 {
736 	struct vhost_net_virtqueue *nvq =
737 		container_of(vq, struct vhost_net_virtqueue, vq);
738 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
739 	struct socket *sock;
740 
741 	sock = vq->private_data;
742 	if (!sock)
743 		return 0;
744 
745 	return vhost_poll_start(poll, sock->file);
746 }
747 
748 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
749 					struct vhost_virtqueue *vq)
750 {
751 	struct socket *sock;
752 
753 	mutex_lock(&vq->mutex);
754 	sock = vq->private_data;
755 	vhost_net_disable_vq(n, vq);
756 	vq->private_data = NULL;
757 	mutex_unlock(&vq->mutex);
758 	return sock;
759 }
760 
761 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
762 			   struct socket **rx_sock)
763 {
764 	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
765 	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
766 }
767 
768 static void vhost_net_flush_vq(struct vhost_net *n, int index)
769 {
770 	vhost_poll_flush(n->poll + index);
771 	vhost_poll_flush(&n->vqs[index].vq.poll);
772 }
773 
774 static void vhost_net_flush(struct vhost_net *n)
775 {
776 	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
777 	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
778 	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
779 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
780 		n->tx_flush = true;
781 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
782 		/* Wait for all lower device DMAs done. */
783 		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
784 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
785 		n->tx_flush = false;
786 		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
787 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
788 	}
789 }
790 
791 static int vhost_net_release(struct inode *inode, struct file *f)
792 {
793 	struct vhost_net *n = f->private_data;
794 	struct socket *tx_sock;
795 	struct socket *rx_sock;
796 
797 	vhost_net_stop(n, &tx_sock, &rx_sock);
798 	vhost_net_flush(n);
799 	vhost_dev_stop(&n->dev);
800 	vhost_dev_cleanup(&n->dev, false);
801 	vhost_net_vq_reset(n);
802 	if (tx_sock)
803 		fput(tx_sock->file);
804 	if (rx_sock)
805 		fput(rx_sock->file);
806 	/* Make sure no callbacks are outstanding */
807 	synchronize_rcu_bh();
808 	/* We do an extra flush before freeing memory,
809 	 * since jobs can re-queue themselves. */
810 	vhost_net_flush(n);
811 	kfree(n->dev.vqs);
812 	kfree(n);
813 	return 0;
814 }
815 
816 static struct socket *get_raw_socket(int fd)
817 {
818 	struct {
819 		struct sockaddr_ll sa;
820 		char  buf[MAX_ADDR_LEN];
821 	} uaddr;
822 	int uaddr_len = sizeof uaddr, r;
823 	struct socket *sock = sockfd_lookup(fd, &r);
824 
825 	if (!sock)
826 		return ERR_PTR(-ENOTSOCK);
827 
828 	/* Parameter checking */
829 	if (sock->sk->sk_type != SOCK_RAW) {
830 		r = -ESOCKTNOSUPPORT;
831 		goto err;
832 	}
833 
834 	r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
835 			       &uaddr_len, 0);
836 	if (r)
837 		goto err;
838 
839 	if (uaddr.sa.sll_family != AF_PACKET) {
840 		r = -EPFNOSUPPORT;
841 		goto err;
842 	}
843 	return sock;
844 err:
845 	fput(sock->file);
846 	return ERR_PTR(r);
847 }
848 
849 static struct socket *get_tap_socket(int fd)
850 {
851 	struct file *file = fget(fd);
852 	struct socket *sock;
853 
854 	if (!file)
855 		return ERR_PTR(-EBADF);
856 	sock = tun_get_socket(file);
857 	if (!IS_ERR(sock))
858 		return sock;
859 	sock = macvtap_get_socket(file);
860 	if (IS_ERR(sock))
861 		fput(file);
862 	return sock;
863 }
864 
865 static struct socket *get_socket(int fd)
866 {
867 	struct socket *sock;
868 
869 	/* special case to disable backend */
870 	if (fd == -1)
871 		return NULL;
872 	sock = get_raw_socket(fd);
873 	if (!IS_ERR(sock))
874 		return sock;
875 	sock = get_tap_socket(fd);
876 	if (!IS_ERR(sock))
877 		return sock;
878 	return ERR_PTR(-ENOTSOCK);
879 }
880 
881 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
882 {
883 	struct socket *sock, *oldsock;
884 	struct vhost_virtqueue *vq;
885 	struct vhost_net_virtqueue *nvq;
886 	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
887 	int r;
888 
889 	mutex_lock(&n->dev.mutex);
890 	r = vhost_dev_check_owner(&n->dev);
891 	if (r)
892 		goto err;
893 
894 	if (index >= VHOST_NET_VQ_MAX) {
895 		r = -ENOBUFS;
896 		goto err;
897 	}
898 	vq = &n->vqs[index].vq;
899 	nvq = &n->vqs[index];
900 	mutex_lock(&vq->mutex);
901 
902 	/* Verify that ring has been setup correctly. */
903 	if (!vhost_vq_access_ok(vq)) {
904 		r = -EFAULT;
905 		goto err_vq;
906 	}
907 	sock = get_socket(fd);
908 	if (IS_ERR(sock)) {
909 		r = PTR_ERR(sock);
910 		goto err_vq;
911 	}
912 
913 	/* start polling new socket */
914 	oldsock = vq->private_data;
915 	if (sock != oldsock) {
916 		ubufs = vhost_net_ubuf_alloc(vq,
917 					     sock && vhost_sock_zcopy(sock));
918 		if (IS_ERR(ubufs)) {
919 			r = PTR_ERR(ubufs);
920 			goto err_ubufs;
921 		}
922 
923 		vhost_net_disable_vq(n, vq);
924 		vq->private_data = sock;
925 		r = vhost_init_used(vq);
926 		if (r)
927 			goto err_used;
928 		r = vhost_net_enable_vq(n, vq);
929 		if (r)
930 			goto err_used;
931 
932 		oldubufs = nvq->ubufs;
933 		nvq->ubufs = ubufs;
934 
935 		n->tx_packets = 0;
936 		n->tx_zcopy_err = 0;
937 		n->tx_flush = false;
938 	}
939 
940 	mutex_unlock(&vq->mutex);
941 
942 	if (oldubufs) {
943 		vhost_net_ubuf_put_wait_and_free(oldubufs);
944 		mutex_lock(&vq->mutex);
945 		vhost_zerocopy_signal_used(n, vq);
946 		mutex_unlock(&vq->mutex);
947 	}
948 
949 	if (oldsock) {
950 		vhost_net_flush_vq(n, index);
951 		fput(oldsock->file);
952 	}
953 
954 	mutex_unlock(&n->dev.mutex);
955 	return 0;
956 
957 err_used:
958 	vq->private_data = oldsock;
959 	vhost_net_enable_vq(n, vq);
960 	if (ubufs)
961 		vhost_net_ubuf_put_wait_and_free(ubufs);
962 err_ubufs:
963 	fput(sock->file);
964 err_vq:
965 	mutex_unlock(&vq->mutex);
966 err:
967 	mutex_unlock(&n->dev.mutex);
968 	return r;
969 }
970 
971 static long vhost_net_reset_owner(struct vhost_net *n)
972 {
973 	struct socket *tx_sock = NULL;
974 	struct socket *rx_sock = NULL;
975 	long err;
976 	struct vhost_memory *memory;
977 
978 	mutex_lock(&n->dev.mutex);
979 	err = vhost_dev_check_owner(&n->dev);
980 	if (err)
981 		goto done;
982 	memory = vhost_dev_reset_owner_prepare();
983 	if (!memory) {
984 		err = -ENOMEM;
985 		goto done;
986 	}
987 	vhost_net_stop(n, &tx_sock, &rx_sock);
988 	vhost_net_flush(n);
989 	vhost_dev_reset_owner(&n->dev, memory);
990 	vhost_net_vq_reset(n);
991 done:
992 	mutex_unlock(&n->dev.mutex);
993 	if (tx_sock)
994 		fput(tx_sock->file);
995 	if (rx_sock)
996 		fput(rx_sock->file);
997 	return err;
998 }
999 
1000 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1001 {
1002 	size_t vhost_hlen, sock_hlen, hdr_len;
1003 	int i;
1004 
1005 	hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1006 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1007 			sizeof(struct virtio_net_hdr);
1008 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1009 		/* vhost provides vnet_hdr */
1010 		vhost_hlen = hdr_len;
1011 		sock_hlen = 0;
1012 	} else {
1013 		/* socket provides vnet_hdr */
1014 		vhost_hlen = 0;
1015 		sock_hlen = hdr_len;
1016 	}
1017 	mutex_lock(&n->dev.mutex);
1018 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1019 	    !vhost_log_access_ok(&n->dev)) {
1020 		mutex_unlock(&n->dev.mutex);
1021 		return -EFAULT;
1022 	}
1023 	n->dev.acked_features = features;
1024 	smp_wmb();
1025 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1026 		mutex_lock(&n->vqs[i].vq.mutex);
1027 		n->vqs[i].vhost_hlen = vhost_hlen;
1028 		n->vqs[i].sock_hlen = sock_hlen;
1029 		mutex_unlock(&n->vqs[i].vq.mutex);
1030 	}
1031 	vhost_net_flush(n);
1032 	mutex_unlock(&n->dev.mutex);
1033 	return 0;
1034 }
1035 
1036 static long vhost_net_set_owner(struct vhost_net *n)
1037 {
1038 	int r;
1039 
1040 	mutex_lock(&n->dev.mutex);
1041 	if (vhost_dev_has_owner(&n->dev)) {
1042 		r = -EBUSY;
1043 		goto out;
1044 	}
1045 	r = vhost_net_set_ubuf_info(n);
1046 	if (r)
1047 		goto out;
1048 	r = vhost_dev_set_owner(&n->dev);
1049 	if (r)
1050 		vhost_net_clear_ubuf_info(n);
1051 	vhost_net_flush(n);
1052 out:
1053 	mutex_unlock(&n->dev.mutex);
1054 	return r;
1055 }
1056 
1057 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1058 			    unsigned long arg)
1059 {
1060 	struct vhost_net *n = f->private_data;
1061 	void __user *argp = (void __user *)arg;
1062 	u64 __user *featurep = argp;
1063 	struct vhost_vring_file backend;
1064 	u64 features;
1065 	int r;
1066 
1067 	switch (ioctl) {
1068 	case VHOST_NET_SET_BACKEND:
1069 		if (copy_from_user(&backend, argp, sizeof backend))
1070 			return -EFAULT;
1071 		return vhost_net_set_backend(n, backend.index, backend.fd);
1072 	case VHOST_GET_FEATURES:
1073 		features = VHOST_NET_FEATURES;
1074 		if (copy_to_user(featurep, &features, sizeof features))
1075 			return -EFAULT;
1076 		return 0;
1077 	case VHOST_SET_FEATURES:
1078 		if (copy_from_user(&features, featurep, sizeof features))
1079 			return -EFAULT;
1080 		if (features & ~VHOST_NET_FEATURES)
1081 			return -EOPNOTSUPP;
1082 		return vhost_net_set_features(n, features);
1083 	case VHOST_RESET_OWNER:
1084 		return vhost_net_reset_owner(n);
1085 	case VHOST_SET_OWNER:
1086 		return vhost_net_set_owner(n);
1087 	default:
1088 		mutex_lock(&n->dev.mutex);
1089 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1090 		if (r == -ENOIOCTLCMD)
1091 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1092 		else
1093 			vhost_net_flush(n);
1094 		mutex_unlock(&n->dev.mutex);
1095 		return r;
1096 	}
1097 }
1098 
1099 #ifdef CONFIG_COMPAT
1100 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1101 				   unsigned long arg)
1102 {
1103 	return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1104 }
1105 #endif
1106 
1107 static const struct file_operations vhost_net_fops = {
1108 	.owner          = THIS_MODULE,
1109 	.release        = vhost_net_release,
1110 	.unlocked_ioctl = vhost_net_ioctl,
1111 #ifdef CONFIG_COMPAT
1112 	.compat_ioctl   = vhost_net_compat_ioctl,
1113 #endif
1114 	.open           = vhost_net_open,
1115 	.llseek		= noop_llseek,
1116 };
1117 
1118 static struct miscdevice vhost_net_misc = {
1119 	.minor = VHOST_NET_MINOR,
1120 	.name = "vhost-net",
1121 	.fops = &vhost_net_fops,
1122 };
1123 
1124 static int vhost_net_init(void)
1125 {
1126 	if (experimental_zcopytx)
1127 		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1128 	return misc_register(&vhost_net_misc);
1129 }
1130 module_init(vhost_net_init);
1131 
1132 static void vhost_net_exit(void)
1133 {
1134 	misc_deregister(&vhost_net_misc);
1135 }
1136 module_exit(vhost_net_exit);
1137 
1138 MODULE_VERSION("0.0.1");
1139 MODULE_LICENSE("GPL v2");
1140 MODULE_AUTHOR("Michael S. Tsirkin");
1141 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1142 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1143 MODULE_ALIAS("devname:vhost-net");
1144