xref: /linux/drivers/net/macvtap.c (revision 2d87650a3bf1b80f7d0d150ee1af3f8a89e5b7aa)
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 static struct macvlan_dev *macvtap_get_vlan_rcu(const struct net_device *dev)
74 {
75 	return rcu_dereference(dev->rx_handler_data);
76 }
77 
78 /*
79  * RCU usage:
80  * The macvtap_queue and the macvlan_dev are loosely coupled, the
81  * pointers from one to the other can only be read while rcu_read_lock
82  * or rtnl is held.
83  *
84  * Both the file and the macvlan_dev hold a reference on the macvtap_queue
85  * through sock_hold(&q->sk). When the macvlan_dev goes away first,
86  * q->vlan becomes inaccessible. When the files gets closed,
87  * macvtap_get_queue() fails.
88  *
89  * There may still be references to the struct sock inside of the
90  * queue from outbound SKBs, but these never reference back to the
91  * file or the dev. The data structure is freed through __sk_free
92  * when both our references and any pending SKBs are gone.
93  */
94 
95 static int macvtap_enable_queue(struct net_device *dev, struct file *file,
96 				struct macvtap_queue *q)
97 {
98 	struct macvlan_dev *vlan = netdev_priv(dev);
99 	int err = -EINVAL;
100 
101 	ASSERT_RTNL();
102 
103 	if (q->enabled)
104 		goto out;
105 
106 	err = 0;
107 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
108 	q->queue_index = vlan->numvtaps;
109 	q->enabled = true;
110 
111 	vlan->numvtaps++;
112 out:
113 	return err;
114 }
115 
116 static int macvtap_set_queue(struct net_device *dev, struct file *file,
117 			     struct macvtap_queue *q)
118 {
119 	struct macvlan_dev *vlan = netdev_priv(dev);
120 	int err = -EBUSY;
121 
122 	rtnl_lock();
123 	if (vlan->numqueues == MAX_MACVTAP_QUEUES)
124 		goto out;
125 
126 	err = 0;
127 	rcu_assign_pointer(q->vlan, vlan);
128 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
129 	sock_hold(&q->sk);
130 
131 	q->file = file;
132 	q->queue_index = vlan->numvtaps;
133 	q->enabled = true;
134 	file->private_data = q;
135 	list_add_tail(&q->next, &vlan->queue_list);
136 
137 	vlan->numvtaps++;
138 	vlan->numqueues++;
139 
140 out:
141 	rtnl_unlock();
142 	return err;
143 }
144 
145 static int macvtap_disable_queue(struct macvtap_queue *q)
146 {
147 	struct macvlan_dev *vlan;
148 	struct macvtap_queue *nq;
149 
150 	ASSERT_RTNL();
151 	if (!q->enabled)
152 		return -EINVAL;
153 
154 	vlan = rtnl_dereference(q->vlan);
155 
156 	if (vlan) {
157 		int index = q->queue_index;
158 		BUG_ON(index >= vlan->numvtaps);
159 		nq = rtnl_dereference(vlan->taps[vlan->numvtaps - 1]);
160 		nq->queue_index = index;
161 
162 		rcu_assign_pointer(vlan->taps[index], nq);
163 		RCU_INIT_POINTER(vlan->taps[vlan->numvtaps - 1], NULL);
164 		q->enabled = false;
165 
166 		vlan->numvtaps--;
167 	}
168 
169 	return 0;
170 }
171 
172 /*
173  * The file owning the queue got closed, give up both
174  * the reference that the files holds as well as the
175  * one from the macvlan_dev if that still exists.
176  *
177  * Using the spinlock makes sure that we don't get
178  * to the queue again after destroying it.
179  */
180 static void macvtap_put_queue(struct macvtap_queue *q)
181 {
182 	struct macvlan_dev *vlan;
183 
184 	rtnl_lock();
185 	vlan = rtnl_dereference(q->vlan);
186 
187 	if (vlan) {
188 		if (q->enabled)
189 			BUG_ON(macvtap_disable_queue(q));
190 
191 		vlan->numqueues--;
192 		RCU_INIT_POINTER(q->vlan, NULL);
193 		sock_put(&q->sk);
194 		list_del_init(&q->next);
195 	}
196 
197 	rtnl_unlock();
198 
199 	synchronize_rcu();
200 	sock_put(&q->sk);
201 }
202 
203 /*
204  * Select a queue based on the rxq of the device on which this packet
205  * arrived. If the incoming device is not mq, calculate a flow hash
206  * to select a queue. If all fails, find the first available queue.
207  * Cache vlan->numvtaps since it can become zero during the execution
208  * of this function.
209  */
210 static struct macvtap_queue *macvtap_get_queue(struct net_device *dev,
211 					       struct sk_buff *skb)
212 {
213 	struct macvlan_dev *vlan = netdev_priv(dev);
214 	struct macvtap_queue *tap = NULL;
215 	/* Access to taps array is protected by rcu, but access to numvtaps
216 	 * isn't. Below we use it to lookup a queue, but treat it as a hint
217 	 * and validate that the result isn't NULL - in case we are
218 	 * racing against queue removal.
219 	 */
220 	int numvtaps = ACCESS_ONCE(vlan->numvtaps);
221 	__u32 rxq;
222 
223 	if (!numvtaps)
224 		goto out;
225 
226 	/* Check if we can use flow to select a queue */
227 	rxq = skb_get_hash(skb);
228 	if (rxq) {
229 		tap = rcu_dereference(vlan->taps[rxq % numvtaps]);
230 		goto out;
231 	}
232 
233 	if (likely(skb_rx_queue_recorded(skb))) {
234 		rxq = skb_get_rx_queue(skb);
235 
236 		while (unlikely(rxq >= numvtaps))
237 			rxq -= numvtaps;
238 
239 		tap = rcu_dereference(vlan->taps[rxq]);
240 		goto out;
241 	}
242 
243 	tap = rcu_dereference(vlan->taps[0]);
244 out:
245 	return tap;
246 }
247 
248 /*
249  * The net_device is going away, give up the reference
250  * that it holds on all queues and safely set the pointer
251  * from the queues to NULL.
252  */
253 static void macvtap_del_queues(struct net_device *dev)
254 {
255 	struct macvlan_dev *vlan = netdev_priv(dev);
256 	struct macvtap_queue *q, *tmp, *qlist[MAX_MACVTAP_QUEUES];
257 	int i, j = 0;
258 
259 	ASSERT_RTNL();
260 	list_for_each_entry_safe(q, tmp, &vlan->queue_list, next) {
261 		list_del_init(&q->next);
262 		qlist[j++] = q;
263 		RCU_INIT_POINTER(q->vlan, NULL);
264 		if (q->enabled)
265 			vlan->numvtaps--;
266 		vlan->numqueues--;
267 	}
268 	for (i = 0; i < vlan->numvtaps; i++)
269 		RCU_INIT_POINTER(vlan->taps[i], NULL);
270 	BUG_ON(vlan->numvtaps);
271 	BUG_ON(vlan->numqueues);
272 	/* guarantee that any future macvtap_set_queue will fail */
273 	vlan->numvtaps = MAX_MACVTAP_QUEUES;
274 
275 	for (--j; j >= 0; j--)
276 		sock_put(&qlist[j]->sk);
277 }
278 
279 static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
280 {
281 	struct sk_buff *skb = *pskb;
282 	struct net_device *dev = skb->dev;
283 	struct macvlan_dev *vlan;
284 	struct macvtap_queue *q;
285 	netdev_features_t features = TAP_FEATURES;
286 
287 	vlan = macvtap_get_vlan_rcu(dev);
288 	if (!vlan)
289 		return RX_HANDLER_PASS;
290 
291 	q = macvtap_get_queue(dev, skb);
292 	if (!q)
293 		return RX_HANDLER_PASS;
294 
295 	if (skb_queue_len(&q->sk.sk_receive_queue) >= dev->tx_queue_len)
296 		goto drop;
297 
298 	skb_push(skb, ETH_HLEN);
299 
300 	/* Apply the forward feature mask so that we perform segmentation
301 	 * according to users wishes.  This only works if VNET_HDR is
302 	 * enabled.
303 	 */
304 	if (q->flags & IFF_VNET_HDR)
305 		features |= vlan->tap_features;
306 	if (netif_needs_gso(skb, features)) {
307 		struct sk_buff *segs = __skb_gso_segment(skb, features, false);
308 
309 		if (IS_ERR(segs))
310 			goto drop;
311 
312 		if (!segs) {
313 			skb_queue_tail(&q->sk.sk_receive_queue, skb);
314 			goto wake_up;
315 		}
316 
317 		kfree_skb(skb);
318 		while (segs) {
319 			struct sk_buff *nskb = segs->next;
320 
321 			segs->next = NULL;
322 			skb_queue_tail(&q->sk.sk_receive_queue, segs);
323 			segs = nskb;
324 		}
325 	} else {
326 		skb_queue_tail(&q->sk.sk_receive_queue, skb);
327 	}
328 
329 wake_up:
330 	wake_up_interruptible_poll(sk_sleep(&q->sk), POLLIN | POLLRDNORM | POLLRDBAND);
331 	return RX_HANDLER_CONSUMED;
332 
333 drop:
334 	/* Count errors/drops only here, thus don't care about args. */
335 	macvlan_count_rx(vlan, 0, 0, 0);
336 	kfree_skb(skb);
337 	return RX_HANDLER_CONSUMED;
338 }
339 
340 static int macvtap_get_minor(struct macvlan_dev *vlan)
341 {
342 	int retval = -ENOMEM;
343 
344 	mutex_lock(&minor_lock);
345 	retval = idr_alloc(&minor_idr, vlan, 1, MACVTAP_NUM_DEVS, GFP_KERNEL);
346 	if (retval >= 0) {
347 		vlan->minor = retval;
348 	} else if (retval == -ENOSPC) {
349 		printk(KERN_ERR "too many macvtap devices\n");
350 		retval = -EINVAL;
351 	}
352 	mutex_unlock(&minor_lock);
353 	return retval < 0 ? retval : 0;
354 }
355 
356 static void macvtap_free_minor(struct macvlan_dev *vlan)
357 {
358 	mutex_lock(&minor_lock);
359 	if (vlan->minor) {
360 		idr_remove(&minor_idr, vlan->minor);
361 		vlan->minor = 0;
362 	}
363 	mutex_unlock(&minor_lock);
364 }
365 
366 static struct net_device *dev_get_by_macvtap_minor(int minor)
367 {
368 	struct net_device *dev = NULL;
369 	struct macvlan_dev *vlan;
370 
371 	mutex_lock(&minor_lock);
372 	vlan = idr_find(&minor_idr, minor);
373 	if (vlan) {
374 		dev = vlan->dev;
375 		dev_hold(dev);
376 	}
377 	mutex_unlock(&minor_lock);
378 	return dev;
379 }
380 
381 static int macvtap_newlink(struct net *src_net,
382 			   struct net_device *dev,
383 			   struct nlattr *tb[],
384 			   struct nlattr *data[])
385 {
386 	struct macvlan_dev *vlan = netdev_priv(dev);
387 	int err;
388 
389 	INIT_LIST_HEAD(&vlan->queue_list);
390 
391 	/* Since macvlan supports all offloads by default, make
392 	 * tap support all offloads also.
393 	 */
394 	vlan->tap_features = TUN_OFFLOADS;
395 
396 	err = netdev_rx_handler_register(dev, macvtap_handle_frame, vlan);
397 	if (err)
398 		return err;
399 
400 	/* Don't put anything that may fail after macvlan_common_newlink
401 	 * because we can't undo what it does.
402 	 */
403 	return macvlan_common_newlink(src_net, dev, tb, data);
404 }
405 
406 static void macvtap_dellink(struct net_device *dev,
407 			    struct list_head *head)
408 {
409 	netdev_rx_handler_unregister(dev);
410 	macvtap_del_queues(dev);
411 	macvlan_dellink(dev, head);
412 }
413 
414 static void macvtap_setup(struct net_device *dev)
415 {
416 	macvlan_common_setup(dev);
417 	dev->tx_queue_len = TUN_READQ_SIZE;
418 }
419 
420 static struct rtnl_link_ops macvtap_link_ops __read_mostly = {
421 	.kind		= "macvtap",
422 	.setup		= macvtap_setup,
423 	.newlink	= macvtap_newlink,
424 	.dellink	= macvtap_dellink,
425 };
426 
427 
428 static void macvtap_sock_write_space(struct sock *sk)
429 {
430 	wait_queue_head_t *wqueue;
431 
432 	if (!sock_writeable(sk) ||
433 	    !test_and_clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags))
434 		return;
435 
436 	wqueue = sk_sleep(sk);
437 	if (wqueue && waitqueue_active(wqueue))
438 		wake_up_interruptible_poll(wqueue, POLLOUT | POLLWRNORM | POLLWRBAND);
439 }
440 
441 static void macvtap_sock_destruct(struct sock *sk)
442 {
443 	skb_queue_purge(&sk->sk_receive_queue);
444 }
445 
446 static int macvtap_open(struct inode *inode, struct file *file)
447 {
448 	struct net *net = current->nsproxy->net_ns;
449 	struct net_device *dev = dev_get_by_macvtap_minor(iminor(inode));
450 	struct macvtap_queue *q;
451 	int err;
452 
453 	err = -ENODEV;
454 	if (!dev)
455 		goto out;
456 
457 	err = -ENOMEM;
458 	q = (struct macvtap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
459 					     &macvtap_proto);
460 	if (!q)
461 		goto out;
462 
463 	RCU_INIT_POINTER(q->sock.wq, &q->wq);
464 	init_waitqueue_head(&q->wq.wait);
465 	q->sock.type = SOCK_RAW;
466 	q->sock.state = SS_CONNECTED;
467 	q->sock.file = file;
468 	q->sock.ops = &macvtap_socket_ops;
469 	sock_init_data(&q->sock, &q->sk);
470 	q->sk.sk_write_space = macvtap_sock_write_space;
471 	q->sk.sk_destruct = macvtap_sock_destruct;
472 	q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
473 	q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
474 
475 	/*
476 	 * so far only KVM virtio_net uses macvtap, enable zero copy between
477 	 * guest kernel and host kernel when lower device supports zerocopy
478 	 *
479 	 * The macvlan supports zerocopy iff the lower device supports zero
480 	 * copy so we don't have to look at the lower device directly.
481 	 */
482 	if ((dev->features & NETIF_F_HIGHDMA) && (dev->features & NETIF_F_SG))
483 		sock_set_flag(&q->sk, SOCK_ZEROCOPY);
484 
485 	err = macvtap_set_queue(dev, file, q);
486 	if (err)
487 		sock_put(&q->sk);
488 
489 out:
490 	if (dev)
491 		dev_put(dev);
492 
493 	return err;
494 }
495 
496 static int macvtap_release(struct inode *inode, struct file *file)
497 {
498 	struct macvtap_queue *q = file->private_data;
499 	macvtap_put_queue(q);
500 	return 0;
501 }
502 
503 static unsigned int macvtap_poll(struct file *file, poll_table * wait)
504 {
505 	struct macvtap_queue *q = file->private_data;
506 	unsigned int mask = POLLERR;
507 
508 	if (!q)
509 		goto out;
510 
511 	mask = 0;
512 	poll_wait(file, &q->wq.wait, wait);
513 
514 	if (!skb_queue_empty(&q->sk.sk_receive_queue))
515 		mask |= POLLIN | POLLRDNORM;
516 
517 	if (sock_writeable(&q->sk) ||
518 	    (!test_and_set_bit(SOCK_ASYNC_NOSPACE, &q->sock.flags) &&
519 	     sock_writeable(&q->sk)))
520 		mask |= POLLOUT | POLLWRNORM;
521 
522 out:
523 	return mask;
524 }
525 
526 static inline struct sk_buff *macvtap_alloc_skb(struct sock *sk, size_t prepad,
527 						size_t len, size_t linear,
528 						int noblock, int *err)
529 {
530 	struct sk_buff *skb;
531 
532 	/* Under a page?  Don't bother with paged skb. */
533 	if (prepad + len < PAGE_SIZE || !linear)
534 		linear = len;
535 
536 	skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
537 				   err, 0);
538 	if (!skb)
539 		return NULL;
540 
541 	skb_reserve(skb, prepad);
542 	skb_put(skb, linear);
543 	skb->data_len = len - linear;
544 	skb->len += len - linear;
545 
546 	return skb;
547 }
548 
549 /*
550  * macvtap_skb_from_vnet_hdr and macvtap_skb_to_vnet_hdr should
551  * be shared with the tun/tap driver.
552  */
553 static int macvtap_skb_from_vnet_hdr(struct sk_buff *skb,
554 				     struct virtio_net_hdr *vnet_hdr)
555 {
556 	unsigned short gso_type = 0;
557 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
558 		switch (vnet_hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
559 		case VIRTIO_NET_HDR_GSO_TCPV4:
560 			gso_type = SKB_GSO_TCPV4;
561 			break;
562 		case VIRTIO_NET_HDR_GSO_TCPV6:
563 			gso_type = SKB_GSO_TCPV6;
564 			break;
565 		case VIRTIO_NET_HDR_GSO_UDP:
566 			gso_type = SKB_GSO_UDP;
567 			break;
568 		default:
569 			return -EINVAL;
570 		}
571 
572 		if (vnet_hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
573 			gso_type |= SKB_GSO_TCP_ECN;
574 
575 		if (vnet_hdr->gso_size == 0)
576 			return -EINVAL;
577 	}
578 
579 	if (vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
580 		if (!skb_partial_csum_set(skb, vnet_hdr->csum_start,
581 					  vnet_hdr->csum_offset))
582 			return -EINVAL;
583 	}
584 
585 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
586 		skb_shinfo(skb)->gso_size = vnet_hdr->gso_size;
587 		skb_shinfo(skb)->gso_type = gso_type;
588 
589 		/* Header must be checked, and gso_segs computed. */
590 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
591 		skb_shinfo(skb)->gso_segs = 0;
592 	}
593 	return 0;
594 }
595 
596 static void macvtap_skb_to_vnet_hdr(const struct sk_buff *skb,
597 				   struct virtio_net_hdr *vnet_hdr)
598 {
599 	memset(vnet_hdr, 0, sizeof(*vnet_hdr));
600 
601 	if (skb_is_gso(skb)) {
602 		struct skb_shared_info *sinfo = skb_shinfo(skb);
603 
604 		/* This is a hint as to how much should be linear. */
605 		vnet_hdr->hdr_len = skb_headlen(skb);
606 		vnet_hdr->gso_size = sinfo->gso_size;
607 		if (sinfo->gso_type & SKB_GSO_TCPV4)
608 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
609 		else if (sinfo->gso_type & SKB_GSO_TCPV6)
610 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
611 		else if (sinfo->gso_type & SKB_GSO_UDP)
612 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
613 		else
614 			BUG();
615 		if (sinfo->gso_type & SKB_GSO_TCP_ECN)
616 			vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
617 	} else
618 		vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
619 
620 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
621 		vnet_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
622 		vnet_hdr->csum_start = skb_checksum_start_offset(skb);
623 		vnet_hdr->csum_offset = skb->csum_offset;
624 	} else if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
625 		vnet_hdr->flags = VIRTIO_NET_HDR_F_DATA_VALID;
626 	} /* else everything is zero */
627 }
628 
629 /* Get packet from user space buffer */
630 static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m,
631 				const struct iovec *iv, unsigned long total_len,
632 				size_t count, int noblock)
633 {
634 	int good_linear = SKB_MAX_HEAD(NET_IP_ALIGN);
635 	struct sk_buff *skb;
636 	struct macvlan_dev *vlan;
637 	unsigned long len = total_len;
638 	int err;
639 	struct virtio_net_hdr vnet_hdr = { 0 };
640 	int vnet_hdr_len = 0;
641 	int copylen = 0;
642 	bool zerocopy = false;
643 	size_t linear;
644 
645 	if (q->flags & IFF_VNET_HDR) {
646 		vnet_hdr_len = q->vnet_hdr_sz;
647 
648 		err = -EINVAL;
649 		if (len < vnet_hdr_len)
650 			goto err;
651 		len -= vnet_hdr_len;
652 
653 		err = memcpy_fromiovecend((void *)&vnet_hdr, iv, 0,
654 					   sizeof(vnet_hdr));
655 		if (err < 0)
656 			goto err;
657 		if ((vnet_hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
658 		     vnet_hdr.csum_start + vnet_hdr.csum_offset + 2 >
659 							vnet_hdr.hdr_len)
660 			vnet_hdr.hdr_len = vnet_hdr.csum_start +
661 						vnet_hdr.csum_offset + 2;
662 		err = -EINVAL;
663 		if (vnet_hdr.hdr_len > len)
664 			goto err;
665 	}
666 
667 	err = -EINVAL;
668 	if (unlikely(len < ETH_HLEN))
669 		goto err;
670 
671 	err = -EMSGSIZE;
672 	if (unlikely(count > UIO_MAXIOV))
673 		goto err;
674 
675 	if (m && m->msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
676 		copylen = vnet_hdr.hdr_len ? vnet_hdr.hdr_len : GOODCOPY_LEN;
677 		if (copylen > good_linear)
678 			copylen = good_linear;
679 		linear = copylen;
680 		if (iov_pages(iv, vnet_hdr_len + copylen, count)
681 		    <= MAX_SKB_FRAGS)
682 			zerocopy = true;
683 	}
684 
685 	if (!zerocopy) {
686 		copylen = len;
687 		if (vnet_hdr.hdr_len > good_linear)
688 			linear = good_linear;
689 		else
690 			linear = vnet_hdr.hdr_len;
691 	}
692 
693 	skb = macvtap_alloc_skb(&q->sk, NET_IP_ALIGN, copylen,
694 				linear, noblock, &err);
695 	if (!skb)
696 		goto err;
697 
698 	if (zerocopy)
699 		err = zerocopy_sg_from_iovec(skb, iv, vnet_hdr_len, count);
700 	else {
701 		err = skb_copy_datagram_from_iovec(skb, 0, iv, vnet_hdr_len,
702 						   len);
703 		if (!err && m && m->msg_control) {
704 			struct ubuf_info *uarg = m->msg_control;
705 			uarg->callback(uarg, false);
706 		}
707 	}
708 
709 	if (err)
710 		goto err_kfree;
711 
712 	skb_set_network_header(skb, ETH_HLEN);
713 	skb_reset_mac_header(skb);
714 	skb->protocol = eth_hdr(skb)->h_proto;
715 
716 	if (vnet_hdr_len) {
717 		err = macvtap_skb_from_vnet_hdr(skb, &vnet_hdr);
718 		if (err)
719 			goto err_kfree;
720 	}
721 
722 	skb_probe_transport_header(skb, ETH_HLEN);
723 
724 	rcu_read_lock();
725 	vlan = rcu_dereference(q->vlan);
726 	/* copy skb_ubuf_info for callback when skb has no error */
727 	if (zerocopy) {
728 		skb_shinfo(skb)->destructor_arg = m->msg_control;
729 		skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
730 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
731 	}
732 	if (vlan) {
733 		skb->dev = vlan->dev;
734 		dev_queue_xmit(skb);
735 	} else {
736 		kfree_skb(skb);
737 	}
738 	rcu_read_unlock();
739 
740 	return total_len;
741 
742 err_kfree:
743 	kfree_skb(skb);
744 
745 err:
746 	rcu_read_lock();
747 	vlan = rcu_dereference(q->vlan);
748 	if (vlan)
749 		this_cpu_inc(vlan->pcpu_stats->tx_dropped);
750 	rcu_read_unlock();
751 
752 	return err;
753 }
754 
755 static ssize_t macvtap_aio_write(struct kiocb *iocb, const struct iovec *iv,
756 				 unsigned long count, loff_t pos)
757 {
758 	struct file *file = iocb->ki_filp;
759 	ssize_t result = -ENOLINK;
760 	struct macvtap_queue *q = file->private_data;
761 
762 	result = macvtap_get_user(q, NULL, iv, iov_length(iv, count), count,
763 				  file->f_flags & O_NONBLOCK);
764 	return result;
765 }
766 
767 /* Put packet to the user space buffer */
768 static ssize_t macvtap_put_user(struct macvtap_queue *q,
769 				const struct sk_buff *skb,
770 				const struct iovec *iv, int len)
771 {
772 	int ret;
773 	int vnet_hdr_len = 0;
774 	int vlan_offset = 0;
775 	int copied, total;
776 
777 	if (q->flags & IFF_VNET_HDR) {
778 		struct virtio_net_hdr vnet_hdr;
779 		vnet_hdr_len = q->vnet_hdr_sz;
780 		if ((len -= vnet_hdr_len) < 0)
781 			return -EINVAL;
782 
783 		macvtap_skb_to_vnet_hdr(skb, &vnet_hdr);
784 
785 		if (memcpy_toiovecend(iv, (void *)&vnet_hdr, 0, sizeof(vnet_hdr)))
786 			return -EFAULT;
787 	}
788 	total = copied = vnet_hdr_len;
789 	total += skb->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 		total += VLAN_HLEN;
805 
806 		copy = min_t(int, vlan_offset, len);
807 		ret = skb_copy_datagram_const_iovec(skb, 0, iv, copied, copy);
808 		len -= copy;
809 		copied += copy;
810 		if (ret || !len)
811 			goto done;
812 
813 		copy = min_t(int, sizeof(veth), len);
814 		ret = memcpy_toiovecend(iv, (void *)&veth, copied, copy);
815 		len -= copy;
816 		copied += copy;
817 		if (ret || !len)
818 			goto done;
819 	}
820 
821 	ret = skb_copy_datagram_const_iovec(skb, vlan_offset, iv, copied, len);
822 
823 done:
824 	return ret ? ret : total;
825 }
826 
827 static ssize_t macvtap_do_read(struct macvtap_queue *q,
828 			       const struct iovec *iv, unsigned long len,
829 			       int noblock)
830 {
831 	DEFINE_WAIT(wait);
832 	struct sk_buff *skb;
833 	ssize_t ret = 0;
834 
835 	while (len) {
836 		if (!noblock)
837 			prepare_to_wait(sk_sleep(&q->sk), &wait,
838 					TASK_INTERRUPTIBLE);
839 
840 		/* Read frames from the queue */
841 		skb = skb_dequeue(&q->sk.sk_receive_queue);
842 		if (!skb) {
843 			if (noblock) {
844 				ret = -EAGAIN;
845 				break;
846 			}
847 			if (signal_pending(current)) {
848 				ret = -ERESTARTSYS;
849 				break;
850 			}
851 			/* Nothing to read, let's sleep */
852 			schedule();
853 			continue;
854 		}
855 		ret = macvtap_put_user(q, skb, iv, len);
856 		kfree_skb(skb);
857 		break;
858 	}
859 
860 	if (!noblock)
861 		finish_wait(sk_sleep(&q->sk), &wait);
862 	return ret;
863 }
864 
865 static ssize_t macvtap_aio_read(struct kiocb *iocb, const struct iovec *iv,
866 				unsigned long count, loff_t pos)
867 {
868 	struct file *file = iocb->ki_filp;
869 	struct macvtap_queue *q = file->private_data;
870 	ssize_t len, ret = 0;
871 
872 	len = iov_length(iv, count);
873 	if (len < 0) {
874 		ret = -EINVAL;
875 		goto out;
876 	}
877 
878 	ret = macvtap_do_read(q, iv, len, file->f_flags & O_NONBLOCK);
879 	ret = min_t(ssize_t, ret, len);
880 	if (ret > 0)
881 		iocb->ki_pos = ret;
882 out:
883 	return ret;
884 }
885 
886 static struct macvlan_dev *macvtap_get_vlan(struct macvtap_queue *q)
887 {
888 	struct macvlan_dev *vlan;
889 
890 	ASSERT_RTNL();
891 	vlan = rtnl_dereference(q->vlan);
892 	if (vlan)
893 		dev_hold(vlan->dev);
894 
895 	return vlan;
896 }
897 
898 static void macvtap_put_vlan(struct macvlan_dev *vlan)
899 {
900 	dev_put(vlan->dev);
901 }
902 
903 static int macvtap_ioctl_set_queue(struct file *file, unsigned int flags)
904 {
905 	struct macvtap_queue *q = file->private_data;
906 	struct macvlan_dev *vlan;
907 	int ret;
908 
909 	vlan = macvtap_get_vlan(q);
910 	if (!vlan)
911 		return -EINVAL;
912 
913 	if (flags & IFF_ATTACH_QUEUE)
914 		ret = macvtap_enable_queue(vlan->dev, file, q);
915 	else if (flags & IFF_DETACH_QUEUE)
916 		ret = macvtap_disable_queue(q);
917 	else
918 		ret = -EINVAL;
919 
920 	macvtap_put_vlan(vlan);
921 	return ret;
922 }
923 
924 static int set_offload(struct macvtap_queue *q, unsigned long arg)
925 {
926 	struct macvlan_dev *vlan;
927 	netdev_features_t features;
928 	netdev_features_t feature_mask = 0;
929 
930 	vlan = rtnl_dereference(q->vlan);
931 	if (!vlan)
932 		return -ENOLINK;
933 
934 	features = vlan->dev->features;
935 
936 	if (arg & TUN_F_CSUM) {
937 		feature_mask = NETIF_F_HW_CSUM;
938 
939 		if (arg & (TUN_F_TSO4 | TUN_F_TSO6)) {
940 			if (arg & TUN_F_TSO_ECN)
941 				feature_mask |= NETIF_F_TSO_ECN;
942 			if (arg & TUN_F_TSO4)
943 				feature_mask |= NETIF_F_TSO;
944 			if (arg & TUN_F_TSO6)
945 				feature_mask |= NETIF_F_TSO6;
946 		}
947 
948 		if (arg & TUN_F_UFO)
949 			feature_mask |= NETIF_F_UFO;
950 	}
951 
952 	/* tun/tap driver inverts the usage for TSO offloads, where
953 	 * setting the TSO bit means that the userspace wants to
954 	 * accept TSO frames and turning it off means that user space
955 	 * does not support TSO.
956 	 * For macvtap, we have to invert it to mean the same thing.
957 	 * When user space turns off TSO, we turn off GSO/LRO so that
958 	 * user-space will not receive TSO frames.
959 	 */
960 	if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6 | NETIF_F_UFO))
961 		features |= RX_OFFLOADS;
962 	else
963 		features &= ~RX_OFFLOADS;
964 
965 	/* tap_features are the same as features on tun/tap and
966 	 * reflect user expectations.
967 	 */
968 	vlan->tap_features = feature_mask;
969 	vlan->set_features = features;
970 	netdev_update_features(vlan->dev);
971 
972 	return 0;
973 }
974 
975 /*
976  * provide compatibility with generic tun/tap interface
977  */
978 static long macvtap_ioctl(struct file *file, unsigned int cmd,
979 			  unsigned long arg)
980 {
981 	struct macvtap_queue *q = file->private_data;
982 	struct macvlan_dev *vlan;
983 	void __user *argp = (void __user *)arg;
984 	struct ifreq __user *ifr = argp;
985 	unsigned int __user *up = argp;
986 	unsigned int u;
987 	int __user *sp = argp;
988 	int s;
989 	int ret;
990 
991 	switch (cmd) {
992 	case TUNSETIFF:
993 		/* ignore the name, just look at flags */
994 		if (get_user(u, &ifr->ifr_flags))
995 			return -EFAULT;
996 
997 		ret = 0;
998 		if ((u & ~(IFF_VNET_HDR | IFF_MULTI_QUEUE)) !=
999 		    (IFF_NO_PI | IFF_TAP))
1000 			ret = -EINVAL;
1001 		else
1002 			q->flags = u;
1003 
1004 		return ret;
1005 
1006 	case TUNGETIFF:
1007 		rtnl_lock();
1008 		vlan = macvtap_get_vlan(q);
1009 		if (!vlan) {
1010 			rtnl_unlock();
1011 			return -ENOLINK;
1012 		}
1013 
1014 		ret = 0;
1015 		if (copy_to_user(&ifr->ifr_name, vlan->dev->name, IFNAMSIZ) ||
1016 		    put_user(q->flags, &ifr->ifr_flags))
1017 			ret = -EFAULT;
1018 		macvtap_put_vlan(vlan);
1019 		rtnl_unlock();
1020 		return ret;
1021 
1022 	case TUNSETQUEUE:
1023 		if (get_user(u, &ifr->ifr_flags))
1024 			return -EFAULT;
1025 		rtnl_lock();
1026 		ret = macvtap_ioctl_set_queue(file, u);
1027 		rtnl_unlock();
1028 		return ret;
1029 
1030 	case TUNGETFEATURES:
1031 		if (put_user(IFF_TAP | IFF_NO_PI | IFF_VNET_HDR |
1032 			     IFF_MULTI_QUEUE, up))
1033 			return -EFAULT;
1034 		return 0;
1035 
1036 	case TUNSETSNDBUF:
1037 		if (get_user(u, up))
1038 			return -EFAULT;
1039 
1040 		q->sk.sk_sndbuf = u;
1041 		return 0;
1042 
1043 	case TUNGETVNETHDRSZ:
1044 		s = q->vnet_hdr_sz;
1045 		if (put_user(s, sp))
1046 			return -EFAULT;
1047 		return 0;
1048 
1049 	case TUNSETVNETHDRSZ:
1050 		if (get_user(s, sp))
1051 			return -EFAULT;
1052 		if (s < (int)sizeof(struct virtio_net_hdr))
1053 			return -EINVAL;
1054 
1055 		q->vnet_hdr_sz = s;
1056 		return 0;
1057 
1058 	case TUNSETOFFLOAD:
1059 		/* let the user check for future flags */
1060 		if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1061 			    TUN_F_TSO_ECN | TUN_F_UFO))
1062 			return -EINVAL;
1063 
1064 		rtnl_lock();
1065 		ret = set_offload(q, arg);
1066 		rtnl_unlock();
1067 		return ret;
1068 
1069 	default:
1070 		return -EINVAL;
1071 	}
1072 }
1073 
1074 #ifdef CONFIG_COMPAT
1075 static long macvtap_compat_ioctl(struct file *file, unsigned int cmd,
1076 				 unsigned long arg)
1077 {
1078 	return macvtap_ioctl(file, cmd, (unsigned long)compat_ptr(arg));
1079 }
1080 #endif
1081 
1082 static const struct file_operations macvtap_fops = {
1083 	.owner		= THIS_MODULE,
1084 	.open		= macvtap_open,
1085 	.release	= macvtap_release,
1086 	.aio_read	= macvtap_aio_read,
1087 	.aio_write	= macvtap_aio_write,
1088 	.poll		= macvtap_poll,
1089 	.llseek		= no_llseek,
1090 	.unlocked_ioctl	= macvtap_ioctl,
1091 #ifdef CONFIG_COMPAT
1092 	.compat_ioctl	= macvtap_compat_ioctl,
1093 #endif
1094 };
1095 
1096 static int macvtap_sendmsg(struct kiocb *iocb, struct socket *sock,
1097 			   struct msghdr *m, size_t total_len)
1098 {
1099 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1100 	return macvtap_get_user(q, m, m->msg_iov, total_len, m->msg_iovlen,
1101 			    m->msg_flags & MSG_DONTWAIT);
1102 }
1103 
1104 static int macvtap_recvmsg(struct kiocb *iocb, struct socket *sock,
1105 			   struct msghdr *m, size_t total_len,
1106 			   int flags)
1107 {
1108 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1109 	int ret;
1110 	if (flags & ~(MSG_DONTWAIT|MSG_TRUNC))
1111 		return -EINVAL;
1112 	ret = macvtap_do_read(q, m->msg_iov, total_len,
1113 			  flags & MSG_DONTWAIT);
1114 	if (ret > total_len) {
1115 		m->msg_flags |= MSG_TRUNC;
1116 		ret = flags & MSG_TRUNC ? ret : total_len;
1117 	}
1118 	return ret;
1119 }
1120 
1121 /* Ops structure to mimic raw sockets with tun */
1122 static const struct proto_ops macvtap_socket_ops = {
1123 	.sendmsg = macvtap_sendmsg,
1124 	.recvmsg = macvtap_recvmsg,
1125 };
1126 
1127 /* Get an underlying socket object from tun file.  Returns error unless file is
1128  * attached to a device.  The returned object works like a packet socket, it
1129  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1130  * holding a reference to the file for as long as the socket is in use. */
1131 struct socket *macvtap_get_socket(struct file *file)
1132 {
1133 	struct macvtap_queue *q;
1134 	if (file->f_op != &macvtap_fops)
1135 		return ERR_PTR(-EINVAL);
1136 	q = file->private_data;
1137 	if (!q)
1138 		return ERR_PTR(-EBADFD);
1139 	return &q->sock;
1140 }
1141 EXPORT_SYMBOL_GPL(macvtap_get_socket);
1142 
1143 static int macvtap_device_event(struct notifier_block *unused,
1144 				unsigned long event, void *ptr)
1145 {
1146 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1147 	struct macvlan_dev *vlan;
1148 	struct device *classdev;
1149 	dev_t devt;
1150 	int err;
1151 
1152 	if (dev->rtnl_link_ops != &macvtap_link_ops)
1153 		return NOTIFY_DONE;
1154 
1155 	vlan = netdev_priv(dev);
1156 
1157 	switch (event) {
1158 	case NETDEV_REGISTER:
1159 		/* Create the device node here after the network device has
1160 		 * been registered but before register_netdevice has
1161 		 * finished running.
1162 		 */
1163 		err = macvtap_get_minor(vlan);
1164 		if (err)
1165 			return notifier_from_errno(err);
1166 
1167 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1168 		classdev = device_create(macvtap_class, &dev->dev, devt,
1169 					 dev, "tap%d", dev->ifindex);
1170 		if (IS_ERR(classdev)) {
1171 			macvtap_free_minor(vlan);
1172 			return notifier_from_errno(PTR_ERR(classdev));
1173 		}
1174 		break;
1175 	case NETDEV_UNREGISTER:
1176 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1177 		device_destroy(macvtap_class, devt);
1178 		macvtap_free_minor(vlan);
1179 		break;
1180 	}
1181 
1182 	return NOTIFY_DONE;
1183 }
1184 
1185 static struct notifier_block macvtap_notifier_block __read_mostly = {
1186 	.notifier_call	= macvtap_device_event,
1187 };
1188 
1189 static int macvtap_init(void)
1190 {
1191 	int err;
1192 
1193 	err = alloc_chrdev_region(&macvtap_major, 0,
1194 				MACVTAP_NUM_DEVS, "macvtap");
1195 	if (err)
1196 		goto out1;
1197 
1198 	cdev_init(&macvtap_cdev, &macvtap_fops);
1199 	err = cdev_add(&macvtap_cdev, macvtap_major, MACVTAP_NUM_DEVS);
1200 	if (err)
1201 		goto out2;
1202 
1203 	macvtap_class = class_create(THIS_MODULE, "macvtap");
1204 	if (IS_ERR(macvtap_class)) {
1205 		err = PTR_ERR(macvtap_class);
1206 		goto out3;
1207 	}
1208 
1209 	err = register_netdevice_notifier(&macvtap_notifier_block);
1210 	if (err)
1211 		goto out4;
1212 
1213 	err = macvlan_link_register(&macvtap_link_ops);
1214 	if (err)
1215 		goto out5;
1216 
1217 	return 0;
1218 
1219 out5:
1220 	unregister_netdevice_notifier(&macvtap_notifier_block);
1221 out4:
1222 	class_unregister(macvtap_class);
1223 out3:
1224 	cdev_del(&macvtap_cdev);
1225 out2:
1226 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1227 out1:
1228 	return err;
1229 }
1230 module_init(macvtap_init);
1231 
1232 static void macvtap_exit(void)
1233 {
1234 	rtnl_link_unregister(&macvtap_link_ops);
1235 	unregister_netdevice_notifier(&macvtap_notifier_block);
1236 	class_unregister(macvtap_class);
1237 	cdev_del(&macvtap_cdev);
1238 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1239 }
1240 module_exit(macvtap_exit);
1241 
1242 MODULE_ALIAS_RTNL_LINK("macvtap");
1243 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1244 MODULE_LICENSE("GPL");
1245