xref: /linux/drivers/vhost/net.c (revision 96c63fa7393d0a346acfe5a91e0c7d4c7782641b)
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 };
66 
67 enum {
68 	VHOST_NET_VQ_RX = 0,
69 	VHOST_NET_VQ_TX = 1,
70 	VHOST_NET_VQ_MAX = 2,
71 };
72 
73 struct vhost_net_ubuf_ref {
74 	/* refcount follows semantics similar to kref:
75 	 *  0: object is released
76 	 *  1: no outstanding ubufs
77 	 * >1: outstanding ubufs
78 	 */
79 	atomic_t refcount;
80 	wait_queue_head_t wait;
81 	struct vhost_virtqueue *vq;
82 };
83 
84 struct vhost_net_virtqueue {
85 	struct vhost_virtqueue vq;
86 	size_t vhost_hlen;
87 	size_t sock_hlen;
88 	/* vhost zerocopy support fields below: */
89 	/* last used idx for outstanding DMA zerocopy buffers */
90 	int upend_idx;
91 	/* first used idx for DMA done zerocopy buffers */
92 	int done_idx;
93 	/* an array of userspace buffers info */
94 	struct ubuf_info *ubuf_info;
95 	/* Reference counting for outstanding ubufs.
96 	 * Protected by vq mutex. Writers must also take device mutex. */
97 	struct vhost_net_ubuf_ref *ubufs;
98 };
99 
100 struct vhost_net {
101 	struct vhost_dev dev;
102 	struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
103 	struct vhost_poll poll[VHOST_NET_VQ_MAX];
104 	/* Number of TX recently submitted.
105 	 * Protected by tx vq lock. */
106 	unsigned tx_packets;
107 	/* Number of times zerocopy TX recently failed.
108 	 * Protected by tx vq lock. */
109 	unsigned tx_zcopy_err;
110 	/* Flush in progress. Protected by tx vq lock. */
111 	bool tx_flush;
112 };
113 
114 static unsigned vhost_net_zcopy_mask __read_mostly;
115 
116 static void vhost_net_enable_zcopy(int vq)
117 {
118 	vhost_net_zcopy_mask |= 0x1 << vq;
119 }
120 
121 static struct vhost_net_ubuf_ref *
122 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
123 {
124 	struct vhost_net_ubuf_ref *ubufs;
125 	/* No zero copy backend? Nothing to count. */
126 	if (!zcopy)
127 		return NULL;
128 	ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
129 	if (!ubufs)
130 		return ERR_PTR(-ENOMEM);
131 	atomic_set(&ubufs->refcount, 1);
132 	init_waitqueue_head(&ubufs->wait);
133 	ubufs->vq = vq;
134 	return ubufs;
135 }
136 
137 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
138 {
139 	int r = atomic_sub_return(1, &ubufs->refcount);
140 	if (unlikely(!r))
141 		wake_up(&ubufs->wait);
142 	return r;
143 }
144 
145 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
146 {
147 	vhost_net_ubuf_put(ubufs);
148 	wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
149 }
150 
151 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
152 {
153 	vhost_net_ubuf_put_and_wait(ubufs);
154 	kfree(ubufs);
155 }
156 
157 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
158 {
159 	int i;
160 
161 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
162 		kfree(n->vqs[i].ubuf_info);
163 		n->vqs[i].ubuf_info = NULL;
164 	}
165 }
166 
167 static int vhost_net_set_ubuf_info(struct vhost_net *n)
168 {
169 	bool zcopy;
170 	int i;
171 
172 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
173 		zcopy = vhost_net_zcopy_mask & (0x1 << i);
174 		if (!zcopy)
175 			continue;
176 		n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
177 					      UIO_MAXIOV, GFP_KERNEL);
178 		if  (!n->vqs[i].ubuf_info)
179 			goto err;
180 	}
181 	return 0;
182 
183 err:
184 	vhost_net_clear_ubuf_info(n);
185 	return -ENOMEM;
186 }
187 
188 static void vhost_net_vq_reset(struct vhost_net *n)
189 {
190 	int i;
191 
192 	vhost_net_clear_ubuf_info(n);
193 
194 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
195 		n->vqs[i].done_idx = 0;
196 		n->vqs[i].upend_idx = 0;
197 		n->vqs[i].ubufs = NULL;
198 		n->vqs[i].vhost_hlen = 0;
199 		n->vqs[i].sock_hlen = 0;
200 	}
201 
202 }
203 
204 static void vhost_net_tx_packet(struct vhost_net *net)
205 {
206 	++net->tx_packets;
207 	if (net->tx_packets < 1024)
208 		return;
209 	net->tx_packets = 0;
210 	net->tx_zcopy_err = 0;
211 }
212 
213 static void vhost_net_tx_err(struct vhost_net *net)
214 {
215 	++net->tx_zcopy_err;
216 }
217 
218 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
219 {
220 	/* TX flush waits for outstanding DMAs to be done.
221 	 * Don't start new DMAs.
222 	 */
223 	return !net->tx_flush &&
224 		net->tx_packets / 64 >= net->tx_zcopy_err;
225 }
226 
227 static bool vhost_sock_zcopy(struct socket *sock)
228 {
229 	return unlikely(experimental_zcopytx) &&
230 		sock_flag(sock->sk, SOCK_ZEROCOPY);
231 }
232 
233 /* In case of DMA done not in order in lower device driver for some reason.
234  * upend_idx is used to track end of used idx, done_idx is used to track head
235  * of used idx. Once lower device DMA done contiguously, we will signal KVM
236  * guest used idx.
237  */
238 static void vhost_zerocopy_signal_used(struct vhost_net *net,
239 				       struct vhost_virtqueue *vq)
240 {
241 	struct vhost_net_virtqueue *nvq =
242 		container_of(vq, struct vhost_net_virtqueue, vq);
243 	int i, add;
244 	int j = 0;
245 
246 	for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
247 		if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
248 			vhost_net_tx_err(net);
249 		if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
250 			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
251 			++j;
252 		} else
253 			break;
254 	}
255 	while (j) {
256 		add = min(UIO_MAXIOV - nvq->done_idx, j);
257 		vhost_add_used_and_signal_n(vq->dev, vq,
258 					    &vq->heads[nvq->done_idx], add);
259 		nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
260 		j -= add;
261 	}
262 }
263 
264 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
265 {
266 	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
267 	struct vhost_virtqueue *vq = ubufs->vq;
268 	int cnt;
269 
270 	rcu_read_lock_bh();
271 
272 	/* set len to mark this desc buffers done DMA */
273 	vq->heads[ubuf->desc].len = success ?
274 		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
275 	cnt = vhost_net_ubuf_put(ubufs);
276 
277 	/*
278 	 * Trigger polling thread if guest stopped submitting new buffers:
279 	 * in this case, the refcount after decrement will eventually reach 1.
280 	 * We also trigger polling periodically after each 16 packets
281 	 * (the value 16 here is more or less arbitrary, it's tuned to trigger
282 	 * less than 10% of times).
283 	 */
284 	if (cnt <= 1 || !(cnt % 16))
285 		vhost_poll_queue(&vq->poll);
286 
287 	rcu_read_unlock_bh();
288 }
289 
290 static inline unsigned long busy_clock(void)
291 {
292 	return local_clock() >> 10;
293 }
294 
295 static bool vhost_can_busy_poll(struct vhost_dev *dev,
296 				unsigned long endtime)
297 {
298 	return likely(!need_resched()) &&
299 	       likely(!time_after(busy_clock(), endtime)) &&
300 	       likely(!signal_pending(current)) &&
301 	       !vhost_has_work(dev);
302 }
303 
304 static void vhost_net_disable_vq(struct vhost_net *n,
305 				 struct vhost_virtqueue *vq)
306 {
307 	struct vhost_net_virtqueue *nvq =
308 		container_of(vq, struct vhost_net_virtqueue, vq);
309 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
310 	if (!vq->private_data)
311 		return;
312 	vhost_poll_stop(poll);
313 }
314 
315 static int vhost_net_enable_vq(struct vhost_net *n,
316 				struct vhost_virtqueue *vq)
317 {
318 	struct vhost_net_virtqueue *nvq =
319 		container_of(vq, struct vhost_net_virtqueue, vq);
320 	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
321 	struct socket *sock;
322 
323 	sock = vq->private_data;
324 	if (!sock)
325 		return 0;
326 
327 	return vhost_poll_start(poll, sock->file);
328 }
329 
330 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
331 				    struct vhost_virtqueue *vq,
332 				    struct iovec iov[], unsigned int iov_size,
333 				    unsigned int *out_num, unsigned int *in_num)
334 {
335 	unsigned long uninitialized_var(endtime);
336 	int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
337 				    out_num, in_num, NULL, NULL);
338 
339 	if (r == vq->num && vq->busyloop_timeout) {
340 		preempt_disable();
341 		endtime = busy_clock() + vq->busyloop_timeout;
342 		while (vhost_can_busy_poll(vq->dev, endtime) &&
343 		       vhost_vq_avail_empty(vq->dev, vq))
344 			cpu_relax_lowlatency();
345 		preempt_enable();
346 		r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
347 					out_num, in_num, NULL, NULL);
348 	}
349 
350 	return r;
351 }
352 
353 /* Expects to be always run from workqueue - which acts as
354  * read-size critical section for our kind of RCU. */
355 static void handle_tx(struct vhost_net *net)
356 {
357 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
358 	struct vhost_virtqueue *vq = &nvq->vq;
359 	unsigned out, in;
360 	int head;
361 	struct msghdr msg = {
362 		.msg_name = NULL,
363 		.msg_namelen = 0,
364 		.msg_control = NULL,
365 		.msg_controllen = 0,
366 		.msg_flags = MSG_DONTWAIT,
367 	};
368 	size_t len, total_len = 0;
369 	int err;
370 	size_t hdr_size;
371 	struct socket *sock;
372 	struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
373 	bool zcopy, zcopy_used;
374 
375 	mutex_lock(&vq->mutex);
376 	sock = vq->private_data;
377 	if (!sock)
378 		goto out;
379 
380 	vhost_disable_notify(&net->dev, vq);
381 
382 	hdr_size = nvq->vhost_hlen;
383 	zcopy = nvq->ubufs;
384 
385 	for (;;) {
386 		/* Release DMAs done buffers first */
387 		if (zcopy)
388 			vhost_zerocopy_signal_used(net, vq);
389 
390 		/* If more outstanding DMAs, queue the work.
391 		 * Handle upend_idx wrap around
392 		 */
393 		if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
394 			      % UIO_MAXIOV == nvq->done_idx))
395 			break;
396 
397 		head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
398 						ARRAY_SIZE(vq->iov),
399 						&out, &in);
400 		/* On error, stop handling until the next kick. */
401 		if (unlikely(head < 0))
402 			break;
403 		/* Nothing new?  Wait for eventfd to tell us they refilled. */
404 		if (head == vq->num) {
405 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
406 				vhost_disable_notify(&net->dev, vq);
407 				continue;
408 			}
409 			break;
410 		}
411 		if (in) {
412 			vq_err(vq, "Unexpected descriptor format for TX: "
413 			       "out %d, int %d\n", out, in);
414 			break;
415 		}
416 		/* Skip header. TODO: support TSO. */
417 		len = iov_length(vq->iov, out);
418 		iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
419 		iov_iter_advance(&msg.msg_iter, hdr_size);
420 		/* Sanity check */
421 		if (!msg_data_left(&msg)) {
422 			vq_err(vq, "Unexpected header len for TX: "
423 			       "%zd expected %zd\n",
424 			       len, hdr_size);
425 			break;
426 		}
427 		len = msg_data_left(&msg);
428 
429 		zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
430 				   && (nvq->upend_idx + 1) % UIO_MAXIOV !=
431 				      nvq->done_idx
432 				   && vhost_net_tx_select_zcopy(net);
433 
434 		/* use msg_control to pass vhost zerocopy ubuf info to skb */
435 		if (zcopy_used) {
436 			struct ubuf_info *ubuf;
437 			ubuf = nvq->ubuf_info + nvq->upend_idx;
438 
439 			vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
440 			vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
441 			ubuf->callback = vhost_zerocopy_callback;
442 			ubuf->ctx = nvq->ubufs;
443 			ubuf->desc = nvq->upend_idx;
444 			msg.msg_control = ubuf;
445 			msg.msg_controllen = sizeof(ubuf);
446 			ubufs = nvq->ubufs;
447 			atomic_inc(&ubufs->refcount);
448 			nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
449 		} else {
450 			msg.msg_control = NULL;
451 			ubufs = NULL;
452 		}
453 		/* TODO: Check specific error and bomb out unless ENOBUFS? */
454 		err = sock->ops->sendmsg(sock, &msg, len);
455 		if (unlikely(err < 0)) {
456 			if (zcopy_used) {
457 				vhost_net_ubuf_put(ubufs);
458 				nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
459 					% UIO_MAXIOV;
460 			}
461 			vhost_discard_vq_desc(vq, 1);
462 			break;
463 		}
464 		if (err != len)
465 			pr_debug("Truncated TX packet: "
466 				 " len %d != %zd\n", err, len);
467 		if (!zcopy_used)
468 			vhost_add_used_and_signal(&net->dev, vq, head, 0);
469 		else
470 			vhost_zerocopy_signal_used(net, vq);
471 		total_len += len;
472 		vhost_net_tx_packet(net);
473 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
474 			vhost_poll_queue(&vq->poll);
475 			break;
476 		}
477 	}
478 out:
479 	mutex_unlock(&vq->mutex);
480 }
481 
482 static int peek_head_len(struct sock *sk)
483 {
484 	struct sk_buff *head;
485 	int len = 0;
486 	unsigned long flags;
487 
488 	spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
489 	head = skb_peek(&sk->sk_receive_queue);
490 	if (likely(head)) {
491 		len = head->len;
492 		if (skb_vlan_tag_present(head))
493 			len += VLAN_HLEN;
494 	}
495 
496 	spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
497 	return len;
498 }
499 
500 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
501 {
502 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
503 	struct vhost_virtqueue *vq = &nvq->vq;
504 	unsigned long uninitialized_var(endtime);
505 	int len = peek_head_len(sk);
506 
507 	if (!len && vq->busyloop_timeout) {
508 		/* Both tx vq and rx socket were polled here */
509 		mutex_lock(&vq->mutex);
510 		vhost_disable_notify(&net->dev, vq);
511 
512 		preempt_disable();
513 		endtime = busy_clock() + vq->busyloop_timeout;
514 
515 		while (vhost_can_busy_poll(&net->dev, endtime) &&
516 		       skb_queue_empty(&sk->sk_receive_queue) &&
517 		       vhost_vq_avail_empty(&net->dev, vq))
518 			cpu_relax_lowlatency();
519 
520 		preempt_enable();
521 
522 		if (vhost_enable_notify(&net->dev, vq))
523 			vhost_poll_queue(&vq->poll);
524 		mutex_unlock(&vq->mutex);
525 
526 		len = peek_head_len(sk);
527 	}
528 
529 	return len;
530 }
531 
532 /* This is a multi-buffer version of vhost_get_desc, that works if
533  *	vq has read descriptors only.
534  * @vq		- the relevant virtqueue
535  * @datalen	- data length we'll be reading
536  * @iovcount	- returned count of io vectors we fill
537  * @log		- vhost log
538  * @log_num	- log offset
539  * @quota       - headcount quota, 1 for big buffer
540  *	returns number of buffer heads allocated, negative on error
541  */
542 static int get_rx_bufs(struct vhost_virtqueue *vq,
543 		       struct vring_used_elem *heads,
544 		       int datalen,
545 		       unsigned *iovcount,
546 		       struct vhost_log *log,
547 		       unsigned *log_num,
548 		       unsigned int quota)
549 {
550 	unsigned int out, in;
551 	int seg = 0;
552 	int headcount = 0;
553 	unsigned d;
554 	int r, nlogs = 0;
555 	/* len is always initialized before use since we are always called with
556 	 * datalen > 0.
557 	 */
558 	u32 uninitialized_var(len);
559 
560 	while (datalen > 0 && headcount < quota) {
561 		if (unlikely(seg >= UIO_MAXIOV)) {
562 			r = -ENOBUFS;
563 			goto err;
564 		}
565 		r = vhost_get_vq_desc(vq, vq->iov + seg,
566 				      ARRAY_SIZE(vq->iov) - seg, &out,
567 				      &in, log, log_num);
568 		if (unlikely(r < 0))
569 			goto err;
570 
571 		d = r;
572 		if (d == vq->num) {
573 			r = 0;
574 			goto err;
575 		}
576 		if (unlikely(out || in <= 0)) {
577 			vq_err(vq, "unexpected descriptor format for RX: "
578 				"out %d, in %d\n", out, in);
579 			r = -EINVAL;
580 			goto err;
581 		}
582 		if (unlikely(log)) {
583 			nlogs += *log_num;
584 			log += *log_num;
585 		}
586 		heads[headcount].id = cpu_to_vhost32(vq, d);
587 		len = iov_length(vq->iov + seg, in);
588 		heads[headcount].len = cpu_to_vhost32(vq, len);
589 		datalen -= len;
590 		++headcount;
591 		seg += in;
592 	}
593 	heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
594 	*iovcount = seg;
595 	if (unlikely(log))
596 		*log_num = nlogs;
597 
598 	/* Detect overrun */
599 	if (unlikely(datalen > 0)) {
600 		r = UIO_MAXIOV + 1;
601 		goto err;
602 	}
603 	return headcount;
604 err:
605 	vhost_discard_vq_desc(vq, headcount);
606 	return r;
607 }
608 
609 /* Expects to be always run from workqueue - which acts as
610  * read-size critical section for our kind of RCU. */
611 static void handle_rx(struct vhost_net *net)
612 {
613 	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
614 	struct vhost_virtqueue *vq = &nvq->vq;
615 	unsigned uninitialized_var(in), log;
616 	struct vhost_log *vq_log;
617 	struct msghdr msg = {
618 		.msg_name = NULL,
619 		.msg_namelen = 0,
620 		.msg_control = NULL, /* FIXME: get and handle RX aux data. */
621 		.msg_controllen = 0,
622 		.msg_flags = MSG_DONTWAIT,
623 	};
624 	struct virtio_net_hdr hdr = {
625 		.flags = 0,
626 		.gso_type = VIRTIO_NET_HDR_GSO_NONE
627 	};
628 	size_t total_len = 0;
629 	int err, mergeable;
630 	s16 headcount;
631 	size_t vhost_hlen, sock_hlen;
632 	size_t vhost_len, sock_len;
633 	struct socket *sock;
634 	struct iov_iter fixup;
635 	__virtio16 num_buffers;
636 
637 	mutex_lock(&vq->mutex);
638 	sock = vq->private_data;
639 	if (!sock)
640 		goto out;
641 	vhost_disable_notify(&net->dev, vq);
642 	vhost_net_disable_vq(net, vq);
643 
644 	vhost_hlen = nvq->vhost_hlen;
645 	sock_hlen = nvq->sock_hlen;
646 
647 	vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
648 		vq->log : NULL;
649 	mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
650 
651 	while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
652 		sock_len += sock_hlen;
653 		vhost_len = sock_len + vhost_hlen;
654 		headcount = get_rx_bufs(vq, vq->heads, vhost_len,
655 					&in, vq_log, &log,
656 					likely(mergeable) ? UIO_MAXIOV : 1);
657 		/* On error, stop handling until the next kick. */
658 		if (unlikely(headcount < 0))
659 			goto out;
660 		/* On overrun, truncate and discard */
661 		if (unlikely(headcount > UIO_MAXIOV)) {
662 			iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
663 			err = sock->ops->recvmsg(sock, &msg,
664 						 1, MSG_DONTWAIT | MSG_TRUNC);
665 			pr_debug("Discarded rx packet: len %zd\n", sock_len);
666 			continue;
667 		}
668 		/* OK, now we need to know about added descriptors. */
669 		if (!headcount) {
670 			if (unlikely(vhost_enable_notify(&net->dev, vq))) {
671 				/* They have slipped one in as we were
672 				 * doing that: check again. */
673 				vhost_disable_notify(&net->dev, vq);
674 				continue;
675 			}
676 			/* Nothing new?  Wait for eventfd to tell us
677 			 * they refilled. */
678 			goto out;
679 		}
680 		/* We don't need to be notified again. */
681 		iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
682 		fixup = msg.msg_iter;
683 		if (unlikely((vhost_hlen))) {
684 			/* We will supply the header ourselves
685 			 * TODO: support TSO.
686 			 */
687 			iov_iter_advance(&msg.msg_iter, vhost_hlen);
688 		}
689 		err = sock->ops->recvmsg(sock, &msg,
690 					 sock_len, MSG_DONTWAIT | MSG_TRUNC);
691 		/* Userspace might have consumed the packet meanwhile:
692 		 * it's not supposed to do this usually, but might be hard
693 		 * to prevent. Discard data we got (if any) and keep going. */
694 		if (unlikely(err != sock_len)) {
695 			pr_debug("Discarded rx packet: "
696 				 " len %d, expected %zd\n", err, sock_len);
697 			vhost_discard_vq_desc(vq, headcount);
698 			continue;
699 		}
700 		/* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
701 		if (unlikely(vhost_hlen)) {
702 			if (copy_to_iter(&hdr, sizeof(hdr),
703 					 &fixup) != sizeof(hdr)) {
704 				vq_err(vq, "Unable to write vnet_hdr "
705 				       "at addr %p\n", vq->iov->iov_base);
706 				goto out;
707 			}
708 		} else {
709 			/* Header came from socket; we'll need to patch
710 			 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
711 			 */
712 			iov_iter_advance(&fixup, sizeof(hdr));
713 		}
714 		/* TODO: Should check and handle checksum. */
715 
716 		num_buffers = cpu_to_vhost16(vq, headcount);
717 		if (likely(mergeable) &&
718 		    copy_to_iter(&num_buffers, sizeof num_buffers,
719 				 &fixup) != sizeof num_buffers) {
720 			vq_err(vq, "Failed num_buffers write");
721 			vhost_discard_vq_desc(vq, headcount);
722 			goto out;
723 		}
724 		vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
725 					    headcount);
726 		if (unlikely(vq_log))
727 			vhost_log_write(vq, vq_log, log, vhost_len);
728 		total_len += vhost_len;
729 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
730 			vhost_poll_queue(&vq->poll);
731 			goto out;
732 		}
733 	}
734 	vhost_net_enable_vq(net, vq);
735 out:
736 	mutex_unlock(&vq->mutex);
737 }
738 
739 static void handle_tx_kick(struct vhost_work *work)
740 {
741 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
742 						  poll.work);
743 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
744 
745 	handle_tx(net);
746 }
747 
748 static void handle_rx_kick(struct vhost_work *work)
749 {
750 	struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
751 						  poll.work);
752 	struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
753 
754 	handle_rx(net);
755 }
756 
757 static void handle_tx_net(struct vhost_work *work)
758 {
759 	struct vhost_net *net = container_of(work, struct vhost_net,
760 					     poll[VHOST_NET_VQ_TX].work);
761 	handle_tx(net);
762 }
763 
764 static void handle_rx_net(struct vhost_work *work)
765 {
766 	struct vhost_net *net = container_of(work, struct vhost_net,
767 					     poll[VHOST_NET_VQ_RX].work);
768 	handle_rx(net);
769 }
770 
771 static int vhost_net_open(struct inode *inode, struct file *f)
772 {
773 	struct vhost_net *n;
774 	struct vhost_dev *dev;
775 	struct vhost_virtqueue **vqs;
776 	int i;
777 
778 	n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
779 	if (!n) {
780 		n = vmalloc(sizeof *n);
781 		if (!n)
782 			return -ENOMEM;
783 	}
784 	vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
785 	if (!vqs) {
786 		kvfree(n);
787 		return -ENOMEM;
788 	}
789 
790 	dev = &n->dev;
791 	vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
792 	vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
793 	n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
794 	n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
795 	for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
796 		n->vqs[i].ubufs = NULL;
797 		n->vqs[i].ubuf_info = NULL;
798 		n->vqs[i].upend_idx = 0;
799 		n->vqs[i].done_idx = 0;
800 		n->vqs[i].vhost_hlen = 0;
801 		n->vqs[i].sock_hlen = 0;
802 	}
803 	vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
804 
805 	vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
806 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
807 
808 	f->private_data = n;
809 
810 	return 0;
811 }
812 
813 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
814 					struct vhost_virtqueue *vq)
815 {
816 	struct socket *sock;
817 
818 	mutex_lock(&vq->mutex);
819 	sock = vq->private_data;
820 	vhost_net_disable_vq(n, vq);
821 	vq->private_data = NULL;
822 	mutex_unlock(&vq->mutex);
823 	return sock;
824 }
825 
826 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
827 			   struct socket **rx_sock)
828 {
829 	*tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
830 	*rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
831 }
832 
833 static void vhost_net_flush_vq(struct vhost_net *n, int index)
834 {
835 	vhost_poll_flush(n->poll + index);
836 	vhost_poll_flush(&n->vqs[index].vq.poll);
837 }
838 
839 static void vhost_net_flush(struct vhost_net *n)
840 {
841 	vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
842 	vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
843 	if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
844 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
845 		n->tx_flush = true;
846 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
847 		/* Wait for all lower device DMAs done. */
848 		vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
849 		mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
850 		n->tx_flush = false;
851 		atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
852 		mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
853 	}
854 }
855 
856 static int vhost_net_release(struct inode *inode, struct file *f)
857 {
858 	struct vhost_net *n = f->private_data;
859 	struct socket *tx_sock;
860 	struct socket *rx_sock;
861 
862 	vhost_net_stop(n, &tx_sock, &rx_sock);
863 	vhost_net_flush(n);
864 	vhost_dev_stop(&n->dev);
865 	vhost_dev_cleanup(&n->dev, false);
866 	vhost_net_vq_reset(n);
867 	if (tx_sock)
868 		sockfd_put(tx_sock);
869 	if (rx_sock)
870 		sockfd_put(rx_sock);
871 	/* Make sure no callbacks are outstanding */
872 	synchronize_rcu_bh();
873 	/* We do an extra flush before freeing memory,
874 	 * since jobs can re-queue themselves. */
875 	vhost_net_flush(n);
876 	kfree(n->dev.vqs);
877 	kvfree(n);
878 	return 0;
879 }
880 
881 static struct socket *get_raw_socket(int fd)
882 {
883 	struct {
884 		struct sockaddr_ll sa;
885 		char  buf[MAX_ADDR_LEN];
886 	} uaddr;
887 	int uaddr_len = sizeof uaddr, r;
888 	struct socket *sock = sockfd_lookup(fd, &r);
889 
890 	if (!sock)
891 		return ERR_PTR(-ENOTSOCK);
892 
893 	/* Parameter checking */
894 	if (sock->sk->sk_type != SOCK_RAW) {
895 		r = -ESOCKTNOSUPPORT;
896 		goto err;
897 	}
898 
899 	r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
900 			       &uaddr_len, 0);
901 	if (r)
902 		goto err;
903 
904 	if (uaddr.sa.sll_family != AF_PACKET) {
905 		r = -EPFNOSUPPORT;
906 		goto err;
907 	}
908 	return sock;
909 err:
910 	sockfd_put(sock);
911 	return ERR_PTR(r);
912 }
913 
914 static struct socket *get_tap_socket(int fd)
915 {
916 	struct file *file = fget(fd);
917 	struct socket *sock;
918 
919 	if (!file)
920 		return ERR_PTR(-EBADF);
921 	sock = tun_get_socket(file);
922 	if (!IS_ERR(sock))
923 		return sock;
924 	sock = macvtap_get_socket(file);
925 	if (IS_ERR(sock))
926 		fput(file);
927 	return sock;
928 }
929 
930 static struct socket *get_socket(int fd)
931 {
932 	struct socket *sock;
933 
934 	/* special case to disable backend */
935 	if (fd == -1)
936 		return NULL;
937 	sock = get_raw_socket(fd);
938 	if (!IS_ERR(sock))
939 		return sock;
940 	sock = get_tap_socket(fd);
941 	if (!IS_ERR(sock))
942 		return sock;
943 	return ERR_PTR(-ENOTSOCK);
944 }
945 
946 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
947 {
948 	struct socket *sock, *oldsock;
949 	struct vhost_virtqueue *vq;
950 	struct vhost_net_virtqueue *nvq;
951 	struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
952 	int r;
953 
954 	mutex_lock(&n->dev.mutex);
955 	r = vhost_dev_check_owner(&n->dev);
956 	if (r)
957 		goto err;
958 
959 	if (index >= VHOST_NET_VQ_MAX) {
960 		r = -ENOBUFS;
961 		goto err;
962 	}
963 	vq = &n->vqs[index].vq;
964 	nvq = &n->vqs[index];
965 	mutex_lock(&vq->mutex);
966 
967 	/* Verify that ring has been setup correctly. */
968 	if (!vhost_vq_access_ok(vq)) {
969 		r = -EFAULT;
970 		goto err_vq;
971 	}
972 	sock = get_socket(fd);
973 	if (IS_ERR(sock)) {
974 		r = PTR_ERR(sock);
975 		goto err_vq;
976 	}
977 
978 	/* start polling new socket */
979 	oldsock = vq->private_data;
980 	if (sock != oldsock) {
981 		ubufs = vhost_net_ubuf_alloc(vq,
982 					     sock && vhost_sock_zcopy(sock));
983 		if (IS_ERR(ubufs)) {
984 			r = PTR_ERR(ubufs);
985 			goto err_ubufs;
986 		}
987 
988 		vhost_net_disable_vq(n, vq);
989 		vq->private_data = sock;
990 		r = vhost_vq_init_access(vq);
991 		if (r)
992 			goto err_used;
993 		r = vhost_net_enable_vq(n, vq);
994 		if (r)
995 			goto err_used;
996 
997 		oldubufs = nvq->ubufs;
998 		nvq->ubufs = ubufs;
999 
1000 		n->tx_packets = 0;
1001 		n->tx_zcopy_err = 0;
1002 		n->tx_flush = false;
1003 	}
1004 
1005 	mutex_unlock(&vq->mutex);
1006 
1007 	if (oldubufs) {
1008 		vhost_net_ubuf_put_wait_and_free(oldubufs);
1009 		mutex_lock(&vq->mutex);
1010 		vhost_zerocopy_signal_used(n, vq);
1011 		mutex_unlock(&vq->mutex);
1012 	}
1013 
1014 	if (oldsock) {
1015 		vhost_net_flush_vq(n, index);
1016 		sockfd_put(oldsock);
1017 	}
1018 
1019 	mutex_unlock(&n->dev.mutex);
1020 	return 0;
1021 
1022 err_used:
1023 	vq->private_data = oldsock;
1024 	vhost_net_enable_vq(n, vq);
1025 	if (ubufs)
1026 		vhost_net_ubuf_put_wait_and_free(ubufs);
1027 err_ubufs:
1028 	sockfd_put(sock);
1029 err_vq:
1030 	mutex_unlock(&vq->mutex);
1031 err:
1032 	mutex_unlock(&n->dev.mutex);
1033 	return r;
1034 }
1035 
1036 static long vhost_net_reset_owner(struct vhost_net *n)
1037 {
1038 	struct socket *tx_sock = NULL;
1039 	struct socket *rx_sock = NULL;
1040 	long err;
1041 	struct vhost_memory *memory;
1042 
1043 	mutex_lock(&n->dev.mutex);
1044 	err = vhost_dev_check_owner(&n->dev);
1045 	if (err)
1046 		goto done;
1047 	memory = vhost_dev_reset_owner_prepare();
1048 	if (!memory) {
1049 		err = -ENOMEM;
1050 		goto done;
1051 	}
1052 	vhost_net_stop(n, &tx_sock, &rx_sock);
1053 	vhost_net_flush(n);
1054 	vhost_dev_reset_owner(&n->dev, memory);
1055 	vhost_net_vq_reset(n);
1056 done:
1057 	mutex_unlock(&n->dev.mutex);
1058 	if (tx_sock)
1059 		sockfd_put(tx_sock);
1060 	if (rx_sock)
1061 		sockfd_put(rx_sock);
1062 	return err;
1063 }
1064 
1065 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1066 {
1067 	size_t vhost_hlen, sock_hlen, hdr_len;
1068 	int i;
1069 
1070 	hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1071 			       (1ULL << VIRTIO_F_VERSION_1))) ?
1072 			sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1073 			sizeof(struct virtio_net_hdr);
1074 	if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1075 		/* vhost provides vnet_hdr */
1076 		vhost_hlen = hdr_len;
1077 		sock_hlen = 0;
1078 	} else {
1079 		/* socket provides vnet_hdr */
1080 		vhost_hlen = 0;
1081 		sock_hlen = hdr_len;
1082 	}
1083 	mutex_lock(&n->dev.mutex);
1084 	if ((features & (1 << VHOST_F_LOG_ALL)) &&
1085 	    !vhost_log_access_ok(&n->dev)) {
1086 		mutex_unlock(&n->dev.mutex);
1087 		return -EFAULT;
1088 	}
1089 	for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1090 		mutex_lock(&n->vqs[i].vq.mutex);
1091 		n->vqs[i].vq.acked_features = features;
1092 		n->vqs[i].vhost_hlen = vhost_hlen;
1093 		n->vqs[i].sock_hlen = sock_hlen;
1094 		mutex_unlock(&n->vqs[i].vq.mutex);
1095 	}
1096 	mutex_unlock(&n->dev.mutex);
1097 	return 0;
1098 }
1099 
1100 static long vhost_net_set_owner(struct vhost_net *n)
1101 {
1102 	int r;
1103 
1104 	mutex_lock(&n->dev.mutex);
1105 	if (vhost_dev_has_owner(&n->dev)) {
1106 		r = -EBUSY;
1107 		goto out;
1108 	}
1109 	r = vhost_net_set_ubuf_info(n);
1110 	if (r)
1111 		goto out;
1112 	r = vhost_dev_set_owner(&n->dev);
1113 	if (r)
1114 		vhost_net_clear_ubuf_info(n);
1115 	vhost_net_flush(n);
1116 out:
1117 	mutex_unlock(&n->dev.mutex);
1118 	return r;
1119 }
1120 
1121 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1122 			    unsigned long arg)
1123 {
1124 	struct vhost_net *n = f->private_data;
1125 	void __user *argp = (void __user *)arg;
1126 	u64 __user *featurep = argp;
1127 	struct vhost_vring_file backend;
1128 	u64 features;
1129 	int r;
1130 
1131 	switch (ioctl) {
1132 	case VHOST_NET_SET_BACKEND:
1133 		if (copy_from_user(&backend, argp, sizeof backend))
1134 			return -EFAULT;
1135 		return vhost_net_set_backend(n, backend.index, backend.fd);
1136 	case VHOST_GET_FEATURES:
1137 		features = VHOST_NET_FEATURES;
1138 		if (copy_to_user(featurep, &features, sizeof features))
1139 			return -EFAULT;
1140 		return 0;
1141 	case VHOST_SET_FEATURES:
1142 		if (copy_from_user(&features, featurep, sizeof features))
1143 			return -EFAULT;
1144 		if (features & ~VHOST_NET_FEATURES)
1145 			return -EOPNOTSUPP;
1146 		return vhost_net_set_features(n, features);
1147 	case VHOST_RESET_OWNER:
1148 		return vhost_net_reset_owner(n);
1149 	case VHOST_SET_OWNER:
1150 		return vhost_net_set_owner(n);
1151 	default:
1152 		mutex_lock(&n->dev.mutex);
1153 		r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1154 		if (r == -ENOIOCTLCMD)
1155 			r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1156 		else
1157 			vhost_net_flush(n);
1158 		mutex_unlock(&n->dev.mutex);
1159 		return r;
1160 	}
1161 }
1162 
1163 #ifdef CONFIG_COMPAT
1164 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1165 				   unsigned long arg)
1166 {
1167 	return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1168 }
1169 #endif
1170 
1171 static const struct file_operations vhost_net_fops = {
1172 	.owner          = THIS_MODULE,
1173 	.release        = vhost_net_release,
1174 	.unlocked_ioctl = vhost_net_ioctl,
1175 #ifdef CONFIG_COMPAT
1176 	.compat_ioctl   = vhost_net_compat_ioctl,
1177 #endif
1178 	.open           = vhost_net_open,
1179 	.llseek		= noop_llseek,
1180 };
1181 
1182 static struct miscdevice vhost_net_misc = {
1183 	.minor = VHOST_NET_MINOR,
1184 	.name = "vhost-net",
1185 	.fops = &vhost_net_fops,
1186 };
1187 
1188 static int vhost_net_init(void)
1189 {
1190 	if (experimental_zcopytx)
1191 		vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1192 	return misc_register(&vhost_net_misc);
1193 }
1194 module_init(vhost_net_init);
1195 
1196 static void vhost_net_exit(void)
1197 {
1198 	misc_deregister(&vhost_net_misc);
1199 }
1200 module_exit(vhost_net_exit);
1201 
1202 MODULE_VERSION("0.0.1");
1203 MODULE_LICENSE("GPL v2");
1204 MODULE_AUTHOR("Michael S. Tsirkin");
1205 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1206 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1207 MODULE_ALIAS("devname:vhost-net");
1208