xref: /openbmc/linux/drivers/net/macvtap.c (revision 79f08d9e)
1 #include <linux/etherdevice.h>
2 #include <linux/if_macvlan.h>
3 #include <linux/if_vlan.h>
4 #include <linux/interrupt.h>
5 #include <linux/nsproxy.h>
6 #include <linux/compat.h>
7 #include <linux/if_tun.h>
8 #include <linux/module.h>
9 #include <linux/skbuff.h>
10 #include <linux/cache.h>
11 #include <linux/sched.h>
12 #include <linux/types.h>
13 #include <linux/slab.h>
14 #include <linux/init.h>
15 #include <linux/wait.h>
16 #include <linux/cdev.h>
17 #include <linux/idr.h>
18 #include <linux/fs.h>
19 
20 #include <net/net_namespace.h>
21 #include <net/rtnetlink.h>
22 #include <net/sock.h>
23 #include <linux/virtio_net.h>
24 
25 /*
26  * A macvtap queue is the central object of this driver, it connects
27  * an open character device to a macvlan interface. There can be
28  * multiple queues on one interface, which map back to queues
29  * implemented in hardware on the underlying device.
30  *
31  * macvtap_proto is used to allocate queues through the sock allocation
32  * mechanism.
33  *
34  */
35 struct macvtap_queue {
36 	struct sock sk;
37 	struct socket sock;
38 	struct socket_wq wq;
39 	int vnet_hdr_sz;
40 	struct macvlan_dev __rcu *vlan;
41 	struct file *file;
42 	unsigned int flags;
43 	u16 queue_index;
44 	bool enabled;
45 	struct list_head next;
46 };
47 
48 static struct proto macvtap_proto = {
49 	.name = "macvtap",
50 	.owner = THIS_MODULE,
51 	.obj_size = sizeof (struct macvtap_queue),
52 };
53 
54 /*
55  * Variables for dealing with macvtaps device numbers.
56  */
57 static dev_t macvtap_major;
58 #define MACVTAP_NUM_DEVS (1U << MINORBITS)
59 static DEFINE_MUTEX(minor_lock);
60 static DEFINE_IDR(minor_idr);
61 
62 #define GOODCOPY_LEN 128
63 static struct class *macvtap_class;
64 static struct cdev macvtap_cdev;
65 
66 static const struct proto_ops macvtap_socket_ops;
67 
68 #define TUN_OFFLOADS (NETIF_F_HW_CSUM | NETIF_F_TSO_ECN | NETIF_F_TSO | \
69 		      NETIF_F_TSO6 | NETIF_F_UFO)
70 #define RX_OFFLOADS (NETIF_F_GRO | NETIF_F_LRO)
71 #define TAP_FEATURES (NETIF_F_GSO | NETIF_F_SG)
72 
73 /*
74  * RCU usage:
75  * The macvtap_queue and the macvlan_dev are loosely coupled, the
76  * pointers from one to the other can only be read while rcu_read_lock
77  * or rtnl is held.
78  *
79  * Both the file and the macvlan_dev hold a reference on the macvtap_queue
80  * through sock_hold(&q->sk). When the macvlan_dev goes away first,
81  * q->vlan becomes inaccessible. When the files gets closed,
82  * macvtap_get_queue() fails.
83  *
84  * There may still be references to the struct sock inside of the
85  * queue from outbound SKBs, but these never reference back to the
86  * file or the dev. The data structure is freed through __sk_free
87  * when both our references and any pending SKBs are gone.
88  */
89 
90 static int macvtap_enable_queue(struct net_device *dev, struct file *file,
91 				struct macvtap_queue *q)
92 {
93 	struct macvlan_dev *vlan = netdev_priv(dev);
94 	int err = -EINVAL;
95 
96 	ASSERT_RTNL();
97 
98 	if (q->enabled)
99 		goto out;
100 
101 	err = 0;
102 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
103 	q->queue_index = vlan->numvtaps;
104 	q->enabled = true;
105 
106 	vlan->numvtaps++;
107 out:
108 	return err;
109 }
110 
111 static int macvtap_set_queue(struct net_device *dev, struct file *file,
112 			     struct macvtap_queue *q)
113 {
114 	struct macvlan_dev *vlan = netdev_priv(dev);
115 	int err = -EBUSY;
116 
117 	rtnl_lock();
118 	if (vlan->numqueues == MAX_MACVTAP_QUEUES)
119 		goto out;
120 
121 	err = 0;
122 	rcu_assign_pointer(q->vlan, vlan);
123 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
124 	sock_hold(&q->sk);
125 
126 	q->file = file;
127 	q->queue_index = vlan->numvtaps;
128 	q->enabled = true;
129 	file->private_data = q;
130 	list_add_tail(&q->next, &vlan->queue_list);
131 
132 	vlan->numvtaps++;
133 	vlan->numqueues++;
134 
135 out:
136 	rtnl_unlock();
137 	return err;
138 }
139 
140 static int macvtap_disable_queue(struct macvtap_queue *q)
141 {
142 	struct macvlan_dev *vlan;
143 	struct macvtap_queue *nq;
144 
145 	ASSERT_RTNL();
146 	if (!q->enabled)
147 		return -EINVAL;
148 
149 	vlan = rtnl_dereference(q->vlan);
150 
151 	if (vlan) {
152 		int index = q->queue_index;
153 		BUG_ON(index >= vlan->numvtaps);
154 		nq = rtnl_dereference(vlan->taps[vlan->numvtaps - 1]);
155 		nq->queue_index = index;
156 
157 		rcu_assign_pointer(vlan->taps[index], nq);
158 		RCU_INIT_POINTER(vlan->taps[vlan->numvtaps - 1], NULL);
159 		q->enabled = false;
160 
161 		vlan->numvtaps--;
162 	}
163 
164 	return 0;
165 }
166 
167 /*
168  * The file owning the queue got closed, give up both
169  * the reference that the files holds as well as the
170  * one from the macvlan_dev if that still exists.
171  *
172  * Using the spinlock makes sure that we don't get
173  * to the queue again after destroying it.
174  */
175 static void macvtap_put_queue(struct macvtap_queue *q)
176 {
177 	struct macvlan_dev *vlan;
178 
179 	rtnl_lock();
180 	vlan = rtnl_dereference(q->vlan);
181 
182 	if (vlan) {
183 		if (q->enabled)
184 			BUG_ON(macvtap_disable_queue(q));
185 
186 		vlan->numqueues--;
187 		RCU_INIT_POINTER(q->vlan, NULL);
188 		sock_put(&q->sk);
189 		list_del_init(&q->next);
190 	}
191 
192 	rtnl_unlock();
193 
194 	synchronize_rcu();
195 	sock_put(&q->sk);
196 }
197 
198 /*
199  * Select a queue based on the rxq of the device on which this packet
200  * arrived. If the incoming device is not mq, calculate a flow hash
201  * to select a queue. If all fails, find the first available queue.
202  * Cache vlan->numvtaps since it can become zero during the execution
203  * of this function.
204  */
205 static struct macvtap_queue *macvtap_get_queue(struct net_device *dev,
206 					       struct sk_buff *skb)
207 {
208 	struct macvlan_dev *vlan = netdev_priv(dev);
209 	struct macvtap_queue *tap = NULL;
210 	/* Access to taps array is protected by rcu, but access to numvtaps
211 	 * isn't. Below we use it to lookup a queue, but treat it as a hint
212 	 * and validate that the result isn't NULL - in case we are
213 	 * racing against queue removal.
214 	 */
215 	int numvtaps = ACCESS_ONCE(vlan->numvtaps);
216 	__u32 rxq;
217 
218 	if (!numvtaps)
219 		goto out;
220 
221 	/* Check if we can use flow to select a queue */
222 	rxq = skb_get_rxhash(skb);
223 	if (rxq) {
224 		tap = rcu_dereference(vlan->taps[rxq % numvtaps]);
225 		goto out;
226 	}
227 
228 	if (likely(skb_rx_queue_recorded(skb))) {
229 		rxq = skb_get_rx_queue(skb);
230 
231 		while (unlikely(rxq >= numvtaps))
232 			rxq -= numvtaps;
233 
234 		tap = rcu_dereference(vlan->taps[rxq]);
235 		goto out;
236 	}
237 
238 	tap = rcu_dereference(vlan->taps[0]);
239 out:
240 	return tap;
241 }
242 
243 /*
244  * The net_device is going away, give up the reference
245  * that it holds on all queues and safely set the pointer
246  * from the queues to NULL.
247  */
248 static void macvtap_del_queues(struct net_device *dev)
249 {
250 	struct macvlan_dev *vlan = netdev_priv(dev);
251 	struct macvtap_queue *q, *tmp, *qlist[MAX_MACVTAP_QUEUES];
252 	int i, j = 0;
253 
254 	ASSERT_RTNL();
255 	list_for_each_entry_safe(q, tmp, &vlan->queue_list, next) {
256 		list_del_init(&q->next);
257 		qlist[j++] = q;
258 		RCU_INIT_POINTER(q->vlan, NULL);
259 		if (q->enabled)
260 			vlan->numvtaps--;
261 		vlan->numqueues--;
262 	}
263 	for (i = 0; i < vlan->numvtaps; i++)
264 		RCU_INIT_POINTER(vlan->taps[i], NULL);
265 	BUG_ON(vlan->numvtaps);
266 	BUG_ON(vlan->numqueues);
267 	/* guarantee that any future macvtap_set_queue will fail */
268 	vlan->numvtaps = MAX_MACVTAP_QUEUES;
269 
270 	for (--j; j >= 0; j--)
271 		sock_put(&qlist[j]->sk);
272 }
273 
274 /*
275  * Forward happens for data that gets sent from one macvlan
276  * endpoint to another one in bridge mode. We just take
277  * the skb and put it into the receive queue.
278  */
279 static int macvtap_forward(struct net_device *dev, struct sk_buff *skb)
280 {
281 	struct macvlan_dev *vlan = netdev_priv(dev);
282 	struct macvtap_queue *q = macvtap_get_queue(dev, skb);
283 	netdev_features_t features = TAP_FEATURES;
284 
285 	if (!q)
286 		goto drop;
287 
288 	if (skb_queue_len(&q->sk.sk_receive_queue) >= dev->tx_queue_len)
289 		goto drop;
290 
291 	skb->dev = dev;
292 	/* Apply the forward feature mask so that we perform segmentation
293 	 * according to users wishes.  This only works if VNET_HDR is
294 	 * enabled.
295 	 */
296 	if (q->flags & IFF_VNET_HDR)
297 		features |= vlan->tap_features;
298 	if (netif_needs_gso(skb, features)) {
299 		struct sk_buff *segs = __skb_gso_segment(skb, features, false);
300 
301 		if (IS_ERR(segs))
302 			goto drop;
303 
304 		if (!segs) {
305 			skb_queue_tail(&q->sk.sk_receive_queue, skb);
306 			goto wake_up;
307 		}
308 
309 		kfree_skb(skb);
310 		while (segs) {
311 			struct sk_buff *nskb = segs->next;
312 
313 			segs->next = NULL;
314 			skb_queue_tail(&q->sk.sk_receive_queue, segs);
315 			segs = nskb;
316 		}
317 	} else {
318 		skb_queue_tail(&q->sk.sk_receive_queue, skb);
319 	}
320 
321 wake_up:
322 	wake_up_interruptible_poll(sk_sleep(&q->sk), POLLIN | POLLRDNORM | POLLRDBAND);
323 	return NET_RX_SUCCESS;
324 
325 drop:
326 	kfree_skb(skb);
327 	return NET_RX_DROP;
328 }
329 
330 /*
331  * Receive is for data from the external interface (lowerdev),
332  * in case of macvtap, we can treat that the same way as
333  * forward, which macvlan cannot.
334  */
335 static int macvtap_receive(struct sk_buff *skb)
336 {
337 	skb_push(skb, ETH_HLEN);
338 	return macvtap_forward(skb->dev, skb);
339 }
340 
341 static int macvtap_get_minor(struct macvlan_dev *vlan)
342 {
343 	int retval = -ENOMEM;
344 
345 	mutex_lock(&minor_lock);
346 	retval = idr_alloc(&minor_idr, vlan, 1, MACVTAP_NUM_DEVS, GFP_KERNEL);
347 	if (retval >= 0) {
348 		vlan->minor = retval;
349 	} else if (retval == -ENOSPC) {
350 		printk(KERN_ERR "too many macvtap devices\n");
351 		retval = -EINVAL;
352 	}
353 	mutex_unlock(&minor_lock);
354 	return retval < 0 ? retval : 0;
355 }
356 
357 static void macvtap_free_minor(struct macvlan_dev *vlan)
358 {
359 	mutex_lock(&minor_lock);
360 	if (vlan->minor) {
361 		idr_remove(&minor_idr, vlan->minor);
362 		vlan->minor = 0;
363 	}
364 	mutex_unlock(&minor_lock);
365 }
366 
367 static struct net_device *dev_get_by_macvtap_minor(int minor)
368 {
369 	struct net_device *dev = NULL;
370 	struct macvlan_dev *vlan;
371 
372 	mutex_lock(&minor_lock);
373 	vlan = idr_find(&minor_idr, minor);
374 	if (vlan) {
375 		dev = vlan->dev;
376 		dev_hold(dev);
377 	}
378 	mutex_unlock(&minor_lock);
379 	return dev;
380 }
381 
382 static int macvtap_newlink(struct net *src_net,
383 			   struct net_device *dev,
384 			   struct nlattr *tb[],
385 			   struct nlattr *data[])
386 {
387 	struct macvlan_dev *vlan = netdev_priv(dev);
388 	INIT_LIST_HEAD(&vlan->queue_list);
389 
390 	/* Since macvlan supports all offloads by default, make
391 	 * tap support all offloads also.
392 	 */
393 	vlan->tap_features = TUN_OFFLOADS;
394 
395 	/* Don't put anything that may fail after macvlan_common_newlink
396 	 * because we can't undo what it does.
397 	 */
398 	return macvlan_common_newlink(src_net, dev, tb, data,
399 				      macvtap_receive, macvtap_forward);
400 }
401 
402 static void macvtap_dellink(struct net_device *dev,
403 			    struct list_head *head)
404 {
405 	macvtap_del_queues(dev);
406 	macvlan_dellink(dev, head);
407 }
408 
409 static void macvtap_setup(struct net_device *dev)
410 {
411 	macvlan_common_setup(dev);
412 	dev->tx_queue_len = TUN_READQ_SIZE;
413 }
414 
415 static struct rtnl_link_ops macvtap_link_ops __read_mostly = {
416 	.kind		= "macvtap",
417 	.setup		= macvtap_setup,
418 	.newlink	= macvtap_newlink,
419 	.dellink	= macvtap_dellink,
420 };
421 
422 
423 static void macvtap_sock_write_space(struct sock *sk)
424 {
425 	wait_queue_head_t *wqueue;
426 
427 	if (!sock_writeable(sk) ||
428 	    !test_and_clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags))
429 		return;
430 
431 	wqueue = sk_sleep(sk);
432 	if (wqueue && waitqueue_active(wqueue))
433 		wake_up_interruptible_poll(wqueue, POLLOUT | POLLWRNORM | POLLWRBAND);
434 }
435 
436 static void macvtap_sock_destruct(struct sock *sk)
437 {
438 	skb_queue_purge(&sk->sk_receive_queue);
439 }
440 
441 static int macvtap_open(struct inode *inode, struct file *file)
442 {
443 	struct net *net = current->nsproxy->net_ns;
444 	struct net_device *dev = dev_get_by_macvtap_minor(iminor(inode));
445 	struct macvtap_queue *q;
446 	int err;
447 
448 	err = -ENODEV;
449 	if (!dev)
450 		goto out;
451 
452 	err = -ENOMEM;
453 	q = (struct macvtap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
454 					     &macvtap_proto);
455 	if (!q)
456 		goto out;
457 
458 	RCU_INIT_POINTER(q->sock.wq, &q->wq);
459 	init_waitqueue_head(&q->wq.wait);
460 	q->sock.type = SOCK_RAW;
461 	q->sock.state = SS_CONNECTED;
462 	q->sock.file = file;
463 	q->sock.ops = &macvtap_socket_ops;
464 	sock_init_data(&q->sock, &q->sk);
465 	q->sk.sk_write_space = macvtap_sock_write_space;
466 	q->sk.sk_destruct = macvtap_sock_destruct;
467 	q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
468 	q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
469 
470 	/*
471 	 * so far only KVM virtio_net uses macvtap, enable zero copy between
472 	 * guest kernel and host kernel when lower device supports zerocopy
473 	 *
474 	 * The macvlan supports zerocopy iff the lower device supports zero
475 	 * copy so we don't have to look at the lower device directly.
476 	 */
477 	if ((dev->features & NETIF_F_HIGHDMA) && (dev->features & NETIF_F_SG))
478 		sock_set_flag(&q->sk, SOCK_ZEROCOPY);
479 
480 	err = macvtap_set_queue(dev, file, q);
481 	if (err)
482 		sock_put(&q->sk);
483 
484 out:
485 	if (dev)
486 		dev_put(dev);
487 
488 	return err;
489 }
490 
491 static int macvtap_release(struct inode *inode, struct file *file)
492 {
493 	struct macvtap_queue *q = file->private_data;
494 	macvtap_put_queue(q);
495 	return 0;
496 }
497 
498 static unsigned int macvtap_poll(struct file *file, poll_table * wait)
499 {
500 	struct macvtap_queue *q = file->private_data;
501 	unsigned int mask = POLLERR;
502 
503 	if (!q)
504 		goto out;
505 
506 	mask = 0;
507 	poll_wait(file, &q->wq.wait, wait);
508 
509 	if (!skb_queue_empty(&q->sk.sk_receive_queue))
510 		mask |= POLLIN | POLLRDNORM;
511 
512 	if (sock_writeable(&q->sk) ||
513 	    (!test_and_set_bit(SOCK_ASYNC_NOSPACE, &q->sock.flags) &&
514 	     sock_writeable(&q->sk)))
515 		mask |= POLLOUT | POLLWRNORM;
516 
517 out:
518 	return mask;
519 }
520 
521 static inline struct sk_buff *macvtap_alloc_skb(struct sock *sk, size_t prepad,
522 						size_t len, size_t linear,
523 						int noblock, int *err)
524 {
525 	struct sk_buff *skb;
526 
527 	/* Under a page?  Don't bother with paged skb. */
528 	if (prepad + len < PAGE_SIZE || !linear)
529 		linear = len;
530 
531 	skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
532 				   err, 0);
533 	if (!skb)
534 		return NULL;
535 
536 	skb_reserve(skb, prepad);
537 	skb_put(skb, linear);
538 	skb->data_len = len - linear;
539 	skb->len += len - linear;
540 
541 	return skb;
542 }
543 
544 /*
545  * macvtap_skb_from_vnet_hdr and macvtap_skb_to_vnet_hdr should
546  * be shared with the tun/tap driver.
547  */
548 static int macvtap_skb_from_vnet_hdr(struct sk_buff *skb,
549 				     struct virtio_net_hdr *vnet_hdr)
550 {
551 	unsigned short gso_type = 0;
552 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
553 		switch (vnet_hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
554 		case VIRTIO_NET_HDR_GSO_TCPV4:
555 			gso_type = SKB_GSO_TCPV4;
556 			break;
557 		case VIRTIO_NET_HDR_GSO_TCPV6:
558 			gso_type = SKB_GSO_TCPV6;
559 			break;
560 		case VIRTIO_NET_HDR_GSO_UDP:
561 			gso_type = SKB_GSO_UDP;
562 			break;
563 		default:
564 			return -EINVAL;
565 		}
566 
567 		if (vnet_hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
568 			gso_type |= SKB_GSO_TCP_ECN;
569 
570 		if (vnet_hdr->gso_size == 0)
571 			return -EINVAL;
572 	}
573 
574 	if (vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
575 		if (!skb_partial_csum_set(skb, vnet_hdr->csum_start,
576 					  vnet_hdr->csum_offset))
577 			return -EINVAL;
578 	}
579 
580 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
581 		skb_shinfo(skb)->gso_size = vnet_hdr->gso_size;
582 		skb_shinfo(skb)->gso_type = gso_type;
583 
584 		/* Header must be checked, and gso_segs computed. */
585 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
586 		skb_shinfo(skb)->gso_segs = 0;
587 	}
588 	return 0;
589 }
590 
591 static int macvtap_skb_to_vnet_hdr(const struct sk_buff *skb,
592 				   struct virtio_net_hdr *vnet_hdr)
593 {
594 	memset(vnet_hdr, 0, sizeof(*vnet_hdr));
595 
596 	if (skb_is_gso(skb)) {
597 		struct skb_shared_info *sinfo = skb_shinfo(skb);
598 
599 		/* This is a hint as to how much should be linear. */
600 		vnet_hdr->hdr_len = skb_headlen(skb);
601 		vnet_hdr->gso_size = sinfo->gso_size;
602 		if (sinfo->gso_type & SKB_GSO_TCPV4)
603 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
604 		else if (sinfo->gso_type & SKB_GSO_TCPV6)
605 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
606 		else if (sinfo->gso_type & SKB_GSO_UDP)
607 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
608 		else
609 			BUG();
610 		if (sinfo->gso_type & SKB_GSO_TCP_ECN)
611 			vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
612 	} else
613 		vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
614 
615 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
616 		vnet_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
617 		vnet_hdr->csum_start = skb_checksum_start_offset(skb);
618 		vnet_hdr->csum_offset = skb->csum_offset;
619 	} else if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
620 		vnet_hdr->flags = VIRTIO_NET_HDR_F_DATA_VALID;
621 	} /* else everything is zero */
622 
623 	return 0;
624 }
625 
626 /* Get packet from user space buffer */
627 static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m,
628 				const struct iovec *iv, unsigned long total_len,
629 				size_t count, int noblock)
630 {
631 	int good_linear = SKB_MAX_HEAD(NET_IP_ALIGN);
632 	struct sk_buff *skb;
633 	struct macvlan_dev *vlan;
634 	unsigned long len = total_len;
635 	int err;
636 	struct virtio_net_hdr vnet_hdr = { 0 };
637 	int vnet_hdr_len = 0;
638 	int copylen = 0;
639 	bool zerocopy = false;
640 	size_t linear;
641 
642 	if (q->flags & IFF_VNET_HDR) {
643 		vnet_hdr_len = q->vnet_hdr_sz;
644 
645 		err = -EINVAL;
646 		if (len < vnet_hdr_len)
647 			goto err;
648 		len -= vnet_hdr_len;
649 
650 		err = memcpy_fromiovecend((void *)&vnet_hdr, iv, 0,
651 					   sizeof(vnet_hdr));
652 		if (err < 0)
653 			goto err;
654 		if ((vnet_hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
655 		     vnet_hdr.csum_start + vnet_hdr.csum_offset + 2 >
656 							vnet_hdr.hdr_len)
657 			vnet_hdr.hdr_len = vnet_hdr.csum_start +
658 						vnet_hdr.csum_offset + 2;
659 		err = -EINVAL;
660 		if (vnet_hdr.hdr_len > len)
661 			goto err;
662 	}
663 
664 	err = -EINVAL;
665 	if (unlikely(len < ETH_HLEN))
666 		goto err;
667 
668 	err = -EMSGSIZE;
669 	if (unlikely(count > UIO_MAXIOV))
670 		goto err;
671 
672 	if (m && m->msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
673 		copylen = vnet_hdr.hdr_len ? vnet_hdr.hdr_len : GOODCOPY_LEN;
674 		if (copylen > good_linear)
675 			copylen = good_linear;
676 		linear = copylen;
677 		if (iov_pages(iv, vnet_hdr_len + copylen, count)
678 		    <= MAX_SKB_FRAGS)
679 			zerocopy = true;
680 	}
681 
682 	if (!zerocopy) {
683 		copylen = len;
684 		if (vnet_hdr.hdr_len > good_linear)
685 			linear = good_linear;
686 		else
687 			linear = vnet_hdr.hdr_len;
688 	}
689 
690 	skb = macvtap_alloc_skb(&q->sk, NET_IP_ALIGN, copylen,
691 				linear, noblock, &err);
692 	if (!skb)
693 		goto err;
694 
695 	if (zerocopy)
696 		err = zerocopy_sg_from_iovec(skb, iv, vnet_hdr_len, count);
697 	else {
698 		err = skb_copy_datagram_from_iovec(skb, 0, iv, vnet_hdr_len,
699 						   len);
700 		if (!err && m && m->msg_control) {
701 			struct ubuf_info *uarg = m->msg_control;
702 			uarg->callback(uarg, false);
703 		}
704 	}
705 
706 	if (err)
707 		goto err_kfree;
708 
709 	skb_set_network_header(skb, ETH_HLEN);
710 	skb_reset_mac_header(skb);
711 	skb->protocol = eth_hdr(skb)->h_proto;
712 
713 	if (vnet_hdr_len) {
714 		err = macvtap_skb_from_vnet_hdr(skb, &vnet_hdr);
715 		if (err)
716 			goto err_kfree;
717 	}
718 
719 	skb_probe_transport_header(skb, ETH_HLEN);
720 
721 	rcu_read_lock();
722 	vlan = rcu_dereference(q->vlan);
723 	/* copy skb_ubuf_info for callback when skb has no error */
724 	if (zerocopy) {
725 		skb_shinfo(skb)->destructor_arg = m->msg_control;
726 		skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
727 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
728 	}
729 	if (vlan) {
730 		local_bh_disable();
731 		macvlan_start_xmit(skb, vlan->dev);
732 		local_bh_enable();
733 	} else {
734 		kfree_skb(skb);
735 	}
736 	rcu_read_unlock();
737 
738 	return total_len;
739 
740 err_kfree:
741 	kfree_skb(skb);
742 
743 err:
744 	rcu_read_lock();
745 	vlan = rcu_dereference(q->vlan);
746 	if (vlan)
747 		vlan->dev->stats.tx_dropped++;
748 	rcu_read_unlock();
749 
750 	return err;
751 }
752 
753 static ssize_t macvtap_aio_write(struct kiocb *iocb, const struct iovec *iv,
754 				 unsigned long count, loff_t pos)
755 {
756 	struct file *file = iocb->ki_filp;
757 	ssize_t result = -ENOLINK;
758 	struct macvtap_queue *q = file->private_data;
759 
760 	result = macvtap_get_user(q, NULL, iv, iov_length(iv, count), count,
761 				  file->f_flags & O_NONBLOCK);
762 	return result;
763 }
764 
765 /* Put packet to the user space buffer */
766 static ssize_t macvtap_put_user(struct macvtap_queue *q,
767 				const struct sk_buff *skb,
768 				const struct iovec *iv, int len)
769 {
770 	struct macvlan_dev *vlan;
771 	int ret;
772 	int vnet_hdr_len = 0;
773 	int vlan_offset = 0;
774 	int copied;
775 
776 	if (q->flags & IFF_VNET_HDR) {
777 		struct virtio_net_hdr vnet_hdr;
778 		vnet_hdr_len = q->vnet_hdr_sz;
779 		if ((len -= vnet_hdr_len) < 0)
780 			return -EINVAL;
781 
782 		ret = macvtap_skb_to_vnet_hdr(skb, &vnet_hdr);
783 		if (ret)
784 			return ret;
785 
786 		if (memcpy_toiovecend(iv, (void *)&vnet_hdr, 0, sizeof(vnet_hdr)))
787 			return -EFAULT;
788 	}
789 	copied = vnet_hdr_len;
790 
791 	if (!vlan_tx_tag_present(skb))
792 		len = min_t(int, skb->len, len);
793 	else {
794 		int copy;
795 		struct {
796 			__be16 h_vlan_proto;
797 			__be16 h_vlan_TCI;
798 		} veth;
799 		veth.h_vlan_proto = skb->vlan_proto;
800 		veth.h_vlan_TCI = htons(vlan_tx_tag_get(skb));
801 
802 		vlan_offset = offsetof(struct vlan_ethhdr, h_vlan_proto);
803 		len = min_t(int, skb->len + VLAN_HLEN, len);
804 
805 		copy = min_t(int, vlan_offset, len);
806 		ret = skb_copy_datagram_const_iovec(skb, 0, iv, copied, copy);
807 		len -= copy;
808 		copied += copy;
809 		if (ret || !len)
810 			goto done;
811 
812 		copy = min_t(int, sizeof(veth), len);
813 		ret = memcpy_toiovecend(iv, (void *)&veth, copied, copy);
814 		len -= copy;
815 		copied += copy;
816 		if (ret || !len)
817 			goto done;
818 	}
819 
820 	ret = skb_copy_datagram_const_iovec(skb, vlan_offset, iv, copied, len);
821 	copied += len;
822 
823 done:
824 	rcu_read_lock();
825 	vlan = rcu_dereference(q->vlan);
826 	if (vlan) {
827 		preempt_disable();
828 		macvlan_count_rx(vlan, copied - vnet_hdr_len, ret == 0, 0);
829 		preempt_enable();
830 	}
831 	rcu_read_unlock();
832 
833 	return ret ? ret : copied;
834 }
835 
836 static ssize_t macvtap_do_read(struct macvtap_queue *q, struct kiocb *iocb,
837 			       const struct iovec *iv, unsigned long len,
838 			       int noblock)
839 {
840 	DEFINE_WAIT(wait);
841 	struct sk_buff *skb;
842 	ssize_t ret = 0;
843 
844 	while (len) {
845 		if (!noblock)
846 			prepare_to_wait(sk_sleep(&q->sk), &wait,
847 					TASK_INTERRUPTIBLE);
848 
849 		/* Read frames from the queue */
850 		skb = skb_dequeue(&q->sk.sk_receive_queue);
851 		if (!skb) {
852 			if (noblock) {
853 				ret = -EAGAIN;
854 				break;
855 			}
856 			if (signal_pending(current)) {
857 				ret = -ERESTARTSYS;
858 				break;
859 			}
860 			/* Nothing to read, let's sleep */
861 			schedule();
862 			continue;
863 		}
864 		ret = macvtap_put_user(q, skb, iv, len);
865 		kfree_skb(skb);
866 		break;
867 	}
868 
869 	if (!noblock)
870 		finish_wait(sk_sleep(&q->sk), &wait);
871 	return ret;
872 }
873 
874 static ssize_t macvtap_aio_read(struct kiocb *iocb, const struct iovec *iv,
875 				unsigned long count, loff_t pos)
876 {
877 	struct file *file = iocb->ki_filp;
878 	struct macvtap_queue *q = file->private_data;
879 	ssize_t len, ret = 0;
880 
881 	len = iov_length(iv, count);
882 	if (len < 0) {
883 		ret = -EINVAL;
884 		goto out;
885 	}
886 
887 	ret = macvtap_do_read(q, iocb, iv, len, file->f_flags & O_NONBLOCK);
888 	ret = min_t(ssize_t, ret, len); /* XXX copied from tun.c. Why? */
889 out:
890 	return ret;
891 }
892 
893 static struct macvlan_dev *macvtap_get_vlan(struct macvtap_queue *q)
894 {
895 	struct macvlan_dev *vlan;
896 
897 	ASSERT_RTNL();
898 	vlan = rtnl_dereference(q->vlan);
899 	if (vlan)
900 		dev_hold(vlan->dev);
901 
902 	return vlan;
903 }
904 
905 static void macvtap_put_vlan(struct macvlan_dev *vlan)
906 {
907 	dev_put(vlan->dev);
908 }
909 
910 static int macvtap_ioctl_set_queue(struct file *file, unsigned int flags)
911 {
912 	struct macvtap_queue *q = file->private_data;
913 	struct macvlan_dev *vlan;
914 	int ret;
915 
916 	vlan = macvtap_get_vlan(q);
917 	if (!vlan)
918 		return -EINVAL;
919 
920 	if (flags & IFF_ATTACH_QUEUE)
921 		ret = macvtap_enable_queue(vlan->dev, file, q);
922 	else if (flags & IFF_DETACH_QUEUE)
923 		ret = macvtap_disable_queue(q);
924 	else
925 		ret = -EINVAL;
926 
927 	macvtap_put_vlan(vlan);
928 	return ret;
929 }
930 
931 static int set_offload(struct macvtap_queue *q, unsigned long arg)
932 {
933 	struct macvlan_dev *vlan;
934 	netdev_features_t features;
935 	netdev_features_t feature_mask = 0;
936 
937 	vlan = rtnl_dereference(q->vlan);
938 	if (!vlan)
939 		return -ENOLINK;
940 
941 	features = vlan->dev->features;
942 
943 	if (arg & TUN_F_CSUM) {
944 		feature_mask = NETIF_F_HW_CSUM;
945 
946 		if (arg & (TUN_F_TSO4 | TUN_F_TSO6)) {
947 			if (arg & TUN_F_TSO_ECN)
948 				feature_mask |= NETIF_F_TSO_ECN;
949 			if (arg & TUN_F_TSO4)
950 				feature_mask |= NETIF_F_TSO;
951 			if (arg & TUN_F_TSO6)
952 				feature_mask |= NETIF_F_TSO6;
953 		}
954 
955 		if (arg & TUN_F_UFO)
956 			feature_mask |= NETIF_F_UFO;
957 	}
958 
959 	/* tun/tap driver inverts the usage for TSO offloads, where
960 	 * setting the TSO bit means that the userspace wants to
961 	 * accept TSO frames and turning it off means that user space
962 	 * does not support TSO.
963 	 * For macvtap, we have to invert it to mean the same thing.
964 	 * When user space turns off TSO, we turn off GSO/LRO so that
965 	 * user-space will not receive TSO frames.
966 	 */
967 	if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6 | NETIF_F_UFO))
968 		features |= RX_OFFLOADS;
969 	else
970 		features &= ~RX_OFFLOADS;
971 
972 	/* tap_features are the same as features on tun/tap and
973 	 * reflect user expectations.
974 	 */
975 	vlan->tap_features = feature_mask;
976 	vlan->set_features = features;
977 	netdev_update_features(vlan->dev);
978 
979 	return 0;
980 }
981 
982 /*
983  * provide compatibility with generic tun/tap interface
984  */
985 static long macvtap_ioctl(struct file *file, unsigned int cmd,
986 			  unsigned long arg)
987 {
988 	struct macvtap_queue *q = file->private_data;
989 	struct macvlan_dev *vlan;
990 	void __user *argp = (void __user *)arg;
991 	struct ifreq __user *ifr = argp;
992 	unsigned int __user *up = argp;
993 	unsigned int u;
994 	int __user *sp = argp;
995 	int s;
996 	int ret;
997 
998 	switch (cmd) {
999 	case TUNSETIFF:
1000 		/* ignore the name, just look at flags */
1001 		if (get_user(u, &ifr->ifr_flags))
1002 			return -EFAULT;
1003 
1004 		ret = 0;
1005 		if ((u & ~(IFF_VNET_HDR | IFF_MULTI_QUEUE)) !=
1006 		    (IFF_NO_PI | IFF_TAP))
1007 			ret = -EINVAL;
1008 		else
1009 			q->flags = u;
1010 
1011 		return ret;
1012 
1013 	case TUNGETIFF:
1014 		rtnl_lock();
1015 		vlan = macvtap_get_vlan(q);
1016 		if (!vlan) {
1017 			rtnl_unlock();
1018 			return -ENOLINK;
1019 		}
1020 
1021 		ret = 0;
1022 		if (copy_to_user(&ifr->ifr_name, vlan->dev->name, IFNAMSIZ) ||
1023 		    put_user(q->flags, &ifr->ifr_flags))
1024 			ret = -EFAULT;
1025 		macvtap_put_vlan(vlan);
1026 		rtnl_unlock();
1027 		return ret;
1028 
1029 	case TUNSETQUEUE:
1030 		if (get_user(u, &ifr->ifr_flags))
1031 			return -EFAULT;
1032 		rtnl_lock();
1033 		ret = macvtap_ioctl_set_queue(file, u);
1034 		rtnl_unlock();
1035 		return ret;
1036 
1037 	case TUNGETFEATURES:
1038 		if (put_user(IFF_TAP | IFF_NO_PI | IFF_VNET_HDR |
1039 			     IFF_MULTI_QUEUE, up))
1040 			return -EFAULT;
1041 		return 0;
1042 
1043 	case TUNSETSNDBUF:
1044 		if (get_user(u, up))
1045 			return -EFAULT;
1046 
1047 		q->sk.sk_sndbuf = u;
1048 		return 0;
1049 
1050 	case TUNGETVNETHDRSZ:
1051 		s = q->vnet_hdr_sz;
1052 		if (put_user(s, sp))
1053 			return -EFAULT;
1054 		return 0;
1055 
1056 	case TUNSETVNETHDRSZ:
1057 		if (get_user(s, sp))
1058 			return -EFAULT;
1059 		if (s < (int)sizeof(struct virtio_net_hdr))
1060 			return -EINVAL;
1061 
1062 		q->vnet_hdr_sz = s;
1063 		return 0;
1064 
1065 	case TUNSETOFFLOAD:
1066 		/* let the user check for future flags */
1067 		if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1068 			    TUN_F_TSO_ECN | TUN_F_UFO))
1069 			return -EINVAL;
1070 
1071 		rtnl_lock();
1072 		ret = set_offload(q, arg);
1073 		rtnl_unlock();
1074 		return ret;
1075 
1076 	default:
1077 		return -EINVAL;
1078 	}
1079 }
1080 
1081 #ifdef CONFIG_COMPAT
1082 static long macvtap_compat_ioctl(struct file *file, unsigned int cmd,
1083 				 unsigned long arg)
1084 {
1085 	return macvtap_ioctl(file, cmd, (unsigned long)compat_ptr(arg));
1086 }
1087 #endif
1088 
1089 static const struct file_operations macvtap_fops = {
1090 	.owner		= THIS_MODULE,
1091 	.open		= macvtap_open,
1092 	.release	= macvtap_release,
1093 	.aio_read	= macvtap_aio_read,
1094 	.aio_write	= macvtap_aio_write,
1095 	.poll		= macvtap_poll,
1096 	.llseek		= no_llseek,
1097 	.unlocked_ioctl	= macvtap_ioctl,
1098 #ifdef CONFIG_COMPAT
1099 	.compat_ioctl	= macvtap_compat_ioctl,
1100 #endif
1101 };
1102 
1103 static int macvtap_sendmsg(struct kiocb *iocb, struct socket *sock,
1104 			   struct msghdr *m, size_t total_len)
1105 {
1106 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1107 	return macvtap_get_user(q, m, m->msg_iov, total_len, m->msg_iovlen,
1108 			    m->msg_flags & MSG_DONTWAIT);
1109 }
1110 
1111 static int macvtap_recvmsg(struct kiocb *iocb, struct socket *sock,
1112 			   struct msghdr *m, size_t total_len,
1113 			   int flags)
1114 {
1115 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1116 	int ret;
1117 	if (flags & ~(MSG_DONTWAIT|MSG_TRUNC))
1118 		return -EINVAL;
1119 	ret = macvtap_do_read(q, iocb, m->msg_iov, total_len,
1120 			  flags & MSG_DONTWAIT);
1121 	if (ret > total_len) {
1122 		m->msg_flags |= MSG_TRUNC;
1123 		ret = flags & MSG_TRUNC ? ret : total_len;
1124 	}
1125 	return ret;
1126 }
1127 
1128 /* Ops structure to mimic raw sockets with tun */
1129 static const struct proto_ops macvtap_socket_ops = {
1130 	.sendmsg = macvtap_sendmsg,
1131 	.recvmsg = macvtap_recvmsg,
1132 };
1133 
1134 /* Get an underlying socket object from tun file.  Returns error unless file is
1135  * attached to a device.  The returned object works like a packet socket, it
1136  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1137  * holding a reference to the file for as long as the socket is in use. */
1138 struct socket *macvtap_get_socket(struct file *file)
1139 {
1140 	struct macvtap_queue *q;
1141 	if (file->f_op != &macvtap_fops)
1142 		return ERR_PTR(-EINVAL);
1143 	q = file->private_data;
1144 	if (!q)
1145 		return ERR_PTR(-EBADFD);
1146 	return &q->sock;
1147 }
1148 EXPORT_SYMBOL_GPL(macvtap_get_socket);
1149 
1150 static int macvtap_device_event(struct notifier_block *unused,
1151 				unsigned long event, void *ptr)
1152 {
1153 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1154 	struct macvlan_dev *vlan;
1155 	struct device *classdev;
1156 	dev_t devt;
1157 	int err;
1158 
1159 	if (dev->rtnl_link_ops != &macvtap_link_ops)
1160 		return NOTIFY_DONE;
1161 
1162 	vlan = netdev_priv(dev);
1163 
1164 	switch (event) {
1165 	case NETDEV_REGISTER:
1166 		/* Create the device node here after the network device has
1167 		 * been registered but before register_netdevice has
1168 		 * finished running.
1169 		 */
1170 		err = macvtap_get_minor(vlan);
1171 		if (err)
1172 			return notifier_from_errno(err);
1173 
1174 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1175 		classdev = device_create(macvtap_class, &dev->dev, devt,
1176 					 dev, "tap%d", dev->ifindex);
1177 		if (IS_ERR(classdev)) {
1178 			macvtap_free_minor(vlan);
1179 			return notifier_from_errno(PTR_ERR(classdev));
1180 		}
1181 		break;
1182 	case NETDEV_UNREGISTER:
1183 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1184 		device_destroy(macvtap_class, devt);
1185 		macvtap_free_minor(vlan);
1186 		break;
1187 	}
1188 
1189 	return NOTIFY_DONE;
1190 }
1191 
1192 static struct notifier_block macvtap_notifier_block __read_mostly = {
1193 	.notifier_call	= macvtap_device_event,
1194 };
1195 
1196 static int macvtap_init(void)
1197 {
1198 	int err;
1199 
1200 	err = alloc_chrdev_region(&macvtap_major, 0,
1201 				MACVTAP_NUM_DEVS, "macvtap");
1202 	if (err)
1203 		goto out1;
1204 
1205 	cdev_init(&macvtap_cdev, &macvtap_fops);
1206 	err = cdev_add(&macvtap_cdev, macvtap_major, MACVTAP_NUM_DEVS);
1207 	if (err)
1208 		goto out2;
1209 
1210 	macvtap_class = class_create(THIS_MODULE, "macvtap");
1211 	if (IS_ERR(macvtap_class)) {
1212 		err = PTR_ERR(macvtap_class);
1213 		goto out3;
1214 	}
1215 
1216 	err = register_netdevice_notifier(&macvtap_notifier_block);
1217 	if (err)
1218 		goto out4;
1219 
1220 	err = macvlan_link_register(&macvtap_link_ops);
1221 	if (err)
1222 		goto out5;
1223 
1224 	return 0;
1225 
1226 out5:
1227 	unregister_netdevice_notifier(&macvtap_notifier_block);
1228 out4:
1229 	class_unregister(macvtap_class);
1230 out3:
1231 	cdev_del(&macvtap_cdev);
1232 out2:
1233 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1234 out1:
1235 	return err;
1236 }
1237 module_init(macvtap_init);
1238 
1239 static void macvtap_exit(void)
1240 {
1241 	rtnl_link_unregister(&macvtap_link_ops);
1242 	unregister_netdevice_notifier(&macvtap_notifier_block);
1243 	class_unregister(macvtap_class);
1244 	cdev_del(&macvtap_cdev);
1245 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1246 }
1247 module_exit(macvtap_exit);
1248 
1249 MODULE_ALIAS_RTNL_LINK("macvtap");
1250 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1251 MODULE_LICENSE("GPL");
1252