xref: /openbmc/linux/drivers/net/macvtap.c (revision 034f90b3)
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/wait.h>
15 #include <linux/cdev.h>
16 #include <linux/idr.h>
17 #include <linux/fs.h>
18 #include <linux/uio.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 #define MACVTAP_FEATURES (IFF_VNET_HDR | IFF_MULTI_QUEUE)
49 
50 #define MACVTAP_VNET_LE 0x80000000
51 
52 static inline u16 macvtap16_to_cpu(struct macvtap_queue *q, __virtio16 val)
53 {
54 	return __virtio16_to_cpu(q->flags & MACVTAP_VNET_LE, val);
55 }
56 
57 static inline __virtio16 cpu_to_macvtap16(struct macvtap_queue *q, u16 val)
58 {
59 	return __cpu_to_virtio16(q->flags & MACVTAP_VNET_LE, val);
60 }
61 
62 static struct proto macvtap_proto = {
63 	.name = "macvtap",
64 	.owner = THIS_MODULE,
65 	.obj_size = sizeof (struct macvtap_queue),
66 };
67 
68 /*
69  * Variables for dealing with macvtaps device numbers.
70  */
71 static dev_t macvtap_major;
72 #define MACVTAP_NUM_DEVS (1U << MINORBITS)
73 static DEFINE_MUTEX(minor_lock);
74 static DEFINE_IDR(minor_idr);
75 
76 #define GOODCOPY_LEN 128
77 static struct class *macvtap_class;
78 static struct cdev macvtap_cdev;
79 
80 static const struct proto_ops macvtap_socket_ops;
81 
82 #define TUN_OFFLOADS (NETIF_F_HW_CSUM | NETIF_F_TSO_ECN | NETIF_F_TSO | \
83 		      NETIF_F_TSO6 | NETIF_F_UFO)
84 #define RX_OFFLOADS (NETIF_F_GRO | NETIF_F_LRO)
85 #define TAP_FEATURES (NETIF_F_GSO | NETIF_F_SG)
86 
87 static struct macvlan_dev *macvtap_get_vlan_rcu(const struct net_device *dev)
88 {
89 	return rcu_dereference(dev->rx_handler_data);
90 }
91 
92 /*
93  * RCU usage:
94  * The macvtap_queue and the macvlan_dev are loosely coupled, the
95  * pointers from one to the other can only be read while rcu_read_lock
96  * or rtnl is held.
97  *
98  * Both the file and the macvlan_dev hold a reference on the macvtap_queue
99  * through sock_hold(&q->sk). When the macvlan_dev goes away first,
100  * q->vlan becomes inaccessible. When the files gets closed,
101  * macvtap_get_queue() fails.
102  *
103  * There may still be references to the struct sock inside of the
104  * queue from outbound SKBs, but these never reference back to the
105  * file or the dev. The data structure is freed through __sk_free
106  * when both our references and any pending SKBs are gone.
107  */
108 
109 static int macvtap_enable_queue(struct net_device *dev, struct file *file,
110 				struct macvtap_queue *q)
111 {
112 	struct macvlan_dev *vlan = netdev_priv(dev);
113 	int err = -EINVAL;
114 
115 	ASSERT_RTNL();
116 
117 	if (q->enabled)
118 		goto out;
119 
120 	err = 0;
121 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
122 	q->queue_index = vlan->numvtaps;
123 	q->enabled = true;
124 
125 	vlan->numvtaps++;
126 out:
127 	return err;
128 }
129 
130 /* Requires RTNL */
131 static int macvtap_set_queue(struct net_device *dev, struct file *file,
132 			     struct macvtap_queue *q)
133 {
134 	struct macvlan_dev *vlan = netdev_priv(dev);
135 
136 	if (vlan->numqueues == MAX_MACVTAP_QUEUES)
137 		return -EBUSY;
138 
139 	rcu_assign_pointer(q->vlan, vlan);
140 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
141 	sock_hold(&q->sk);
142 
143 	q->file = file;
144 	q->queue_index = vlan->numvtaps;
145 	q->enabled = true;
146 	file->private_data = q;
147 	list_add_tail(&q->next, &vlan->queue_list);
148 
149 	vlan->numvtaps++;
150 	vlan->numqueues++;
151 
152 	return 0;
153 }
154 
155 static int macvtap_disable_queue(struct macvtap_queue *q)
156 {
157 	struct macvlan_dev *vlan;
158 	struct macvtap_queue *nq;
159 
160 	ASSERT_RTNL();
161 	if (!q->enabled)
162 		return -EINVAL;
163 
164 	vlan = rtnl_dereference(q->vlan);
165 
166 	if (vlan) {
167 		int index = q->queue_index;
168 		BUG_ON(index >= vlan->numvtaps);
169 		nq = rtnl_dereference(vlan->taps[vlan->numvtaps - 1]);
170 		nq->queue_index = index;
171 
172 		rcu_assign_pointer(vlan->taps[index], nq);
173 		RCU_INIT_POINTER(vlan->taps[vlan->numvtaps - 1], NULL);
174 		q->enabled = false;
175 
176 		vlan->numvtaps--;
177 	}
178 
179 	return 0;
180 }
181 
182 /*
183  * The file owning the queue got closed, give up both
184  * the reference that the files holds as well as the
185  * one from the macvlan_dev if that still exists.
186  *
187  * Using the spinlock makes sure that we don't get
188  * to the queue again after destroying it.
189  */
190 static void macvtap_put_queue(struct macvtap_queue *q)
191 {
192 	struct macvlan_dev *vlan;
193 
194 	rtnl_lock();
195 	vlan = rtnl_dereference(q->vlan);
196 
197 	if (vlan) {
198 		if (q->enabled)
199 			BUG_ON(macvtap_disable_queue(q));
200 
201 		vlan->numqueues--;
202 		RCU_INIT_POINTER(q->vlan, NULL);
203 		sock_put(&q->sk);
204 		list_del_init(&q->next);
205 	}
206 
207 	rtnl_unlock();
208 
209 	synchronize_rcu();
210 	sock_put(&q->sk);
211 }
212 
213 /*
214  * Select a queue based on the rxq of the device on which this packet
215  * arrived. If the incoming device is not mq, calculate a flow hash
216  * to select a queue. If all fails, find the first available queue.
217  * Cache vlan->numvtaps since it can become zero during the execution
218  * of this function.
219  */
220 static struct macvtap_queue *macvtap_get_queue(struct net_device *dev,
221 					       struct sk_buff *skb)
222 {
223 	struct macvlan_dev *vlan = netdev_priv(dev);
224 	struct macvtap_queue *tap = NULL;
225 	/* Access to taps array is protected by rcu, but access to numvtaps
226 	 * isn't. Below we use it to lookup a queue, but treat it as a hint
227 	 * and validate that the result isn't NULL - in case we are
228 	 * racing against queue removal.
229 	 */
230 	int numvtaps = ACCESS_ONCE(vlan->numvtaps);
231 	__u32 rxq;
232 
233 	if (!numvtaps)
234 		goto out;
235 
236 	/* Check if we can use flow to select a queue */
237 	rxq = skb_get_hash(skb);
238 	if (rxq) {
239 		tap = rcu_dereference(vlan->taps[rxq % numvtaps]);
240 		goto out;
241 	}
242 
243 	if (likely(skb_rx_queue_recorded(skb))) {
244 		rxq = skb_get_rx_queue(skb);
245 
246 		while (unlikely(rxq >= numvtaps))
247 			rxq -= numvtaps;
248 
249 		tap = rcu_dereference(vlan->taps[rxq]);
250 		goto out;
251 	}
252 
253 	tap = rcu_dereference(vlan->taps[0]);
254 out:
255 	return tap;
256 }
257 
258 /*
259  * The net_device is going away, give up the reference
260  * that it holds on all queues and safely set the pointer
261  * from the queues to NULL.
262  */
263 static void macvtap_del_queues(struct net_device *dev)
264 {
265 	struct macvlan_dev *vlan = netdev_priv(dev);
266 	struct macvtap_queue *q, *tmp, *qlist[MAX_MACVTAP_QUEUES];
267 	int i, j = 0;
268 
269 	ASSERT_RTNL();
270 	list_for_each_entry_safe(q, tmp, &vlan->queue_list, next) {
271 		list_del_init(&q->next);
272 		qlist[j++] = q;
273 		RCU_INIT_POINTER(q->vlan, NULL);
274 		if (q->enabled)
275 			vlan->numvtaps--;
276 		vlan->numqueues--;
277 	}
278 	for (i = 0; i < vlan->numvtaps; i++)
279 		RCU_INIT_POINTER(vlan->taps[i], NULL);
280 	BUG_ON(vlan->numvtaps);
281 	BUG_ON(vlan->numqueues);
282 	/* guarantee that any future macvtap_set_queue will fail */
283 	vlan->numvtaps = MAX_MACVTAP_QUEUES;
284 
285 	for (--j; j >= 0; j--)
286 		sock_put(&qlist[j]->sk);
287 }
288 
289 static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
290 {
291 	struct sk_buff *skb = *pskb;
292 	struct net_device *dev = skb->dev;
293 	struct macvlan_dev *vlan;
294 	struct macvtap_queue *q;
295 	netdev_features_t features = TAP_FEATURES;
296 
297 	vlan = macvtap_get_vlan_rcu(dev);
298 	if (!vlan)
299 		return RX_HANDLER_PASS;
300 
301 	q = macvtap_get_queue(dev, skb);
302 	if (!q)
303 		return RX_HANDLER_PASS;
304 
305 	if (skb_queue_len(&q->sk.sk_receive_queue) >= dev->tx_queue_len)
306 		goto drop;
307 
308 	skb_push(skb, ETH_HLEN);
309 
310 	/* Apply the forward feature mask so that we perform segmentation
311 	 * according to users wishes.  This only works if VNET_HDR is
312 	 * enabled.
313 	 */
314 	if (q->flags & IFF_VNET_HDR)
315 		features |= vlan->tap_features;
316 	if (netif_needs_gso(dev, skb, features)) {
317 		struct sk_buff *segs = __skb_gso_segment(skb, features, false);
318 
319 		if (IS_ERR(segs))
320 			goto drop;
321 
322 		if (!segs) {
323 			skb_queue_tail(&q->sk.sk_receive_queue, skb);
324 			goto wake_up;
325 		}
326 
327 		kfree_skb(skb);
328 		while (segs) {
329 			struct sk_buff *nskb = segs->next;
330 
331 			segs->next = NULL;
332 			skb_queue_tail(&q->sk.sk_receive_queue, segs);
333 			segs = nskb;
334 		}
335 	} else {
336 		/* If we receive a partial checksum and the tap side
337 		 * doesn't support checksum offload, compute the checksum.
338 		 * Note: it doesn't matter which checksum feature to
339 		 *        check, we either support them all or none.
340 		 */
341 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
342 		    !(features & NETIF_F_ALL_CSUM) &&
343 		    skb_checksum_help(skb))
344 			goto drop;
345 		skb_queue_tail(&q->sk.sk_receive_queue, skb);
346 	}
347 
348 wake_up:
349 	wake_up_interruptible_poll(sk_sleep(&q->sk), POLLIN | POLLRDNORM | POLLRDBAND);
350 	return RX_HANDLER_CONSUMED;
351 
352 drop:
353 	/* Count errors/drops only here, thus don't care about args. */
354 	macvlan_count_rx(vlan, 0, 0, 0);
355 	kfree_skb(skb);
356 	return RX_HANDLER_CONSUMED;
357 }
358 
359 static int macvtap_get_minor(struct macvlan_dev *vlan)
360 {
361 	int retval = -ENOMEM;
362 
363 	mutex_lock(&minor_lock);
364 	retval = idr_alloc(&minor_idr, vlan, 1, MACVTAP_NUM_DEVS, GFP_KERNEL);
365 	if (retval >= 0) {
366 		vlan->minor = retval;
367 	} else if (retval == -ENOSPC) {
368 		printk(KERN_ERR "too many macvtap devices\n");
369 		retval = -EINVAL;
370 	}
371 	mutex_unlock(&minor_lock);
372 	return retval < 0 ? retval : 0;
373 }
374 
375 static void macvtap_free_minor(struct macvlan_dev *vlan)
376 {
377 	mutex_lock(&minor_lock);
378 	if (vlan->minor) {
379 		idr_remove(&minor_idr, vlan->minor);
380 		vlan->minor = 0;
381 	}
382 	mutex_unlock(&minor_lock);
383 }
384 
385 static struct net_device *dev_get_by_macvtap_minor(int minor)
386 {
387 	struct net_device *dev = NULL;
388 	struct macvlan_dev *vlan;
389 
390 	mutex_lock(&minor_lock);
391 	vlan = idr_find(&minor_idr, minor);
392 	if (vlan) {
393 		dev = vlan->dev;
394 		dev_hold(dev);
395 	}
396 	mutex_unlock(&minor_lock);
397 	return dev;
398 }
399 
400 static int macvtap_newlink(struct net *src_net,
401 			   struct net_device *dev,
402 			   struct nlattr *tb[],
403 			   struct nlattr *data[])
404 {
405 	struct macvlan_dev *vlan = netdev_priv(dev);
406 	int err;
407 
408 	INIT_LIST_HEAD(&vlan->queue_list);
409 
410 	/* Since macvlan supports all offloads by default, make
411 	 * tap support all offloads also.
412 	 */
413 	vlan->tap_features = TUN_OFFLOADS;
414 
415 	err = netdev_rx_handler_register(dev, macvtap_handle_frame, vlan);
416 	if (err)
417 		return err;
418 
419 	/* Don't put anything that may fail after macvlan_common_newlink
420 	 * because we can't undo what it does.
421 	 */
422 	return macvlan_common_newlink(src_net, dev, tb, data);
423 }
424 
425 static void macvtap_dellink(struct net_device *dev,
426 			    struct list_head *head)
427 {
428 	netdev_rx_handler_unregister(dev);
429 	macvtap_del_queues(dev);
430 	macvlan_dellink(dev, head);
431 }
432 
433 static void macvtap_setup(struct net_device *dev)
434 {
435 	macvlan_common_setup(dev);
436 	dev->tx_queue_len = TUN_READQ_SIZE;
437 }
438 
439 static struct rtnl_link_ops macvtap_link_ops __read_mostly = {
440 	.kind		= "macvtap",
441 	.setup		= macvtap_setup,
442 	.newlink	= macvtap_newlink,
443 	.dellink	= macvtap_dellink,
444 };
445 
446 
447 static void macvtap_sock_write_space(struct sock *sk)
448 {
449 	wait_queue_head_t *wqueue;
450 
451 	if (!sock_writeable(sk) ||
452 	    !test_and_clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags))
453 		return;
454 
455 	wqueue = sk_sleep(sk);
456 	if (wqueue && waitqueue_active(wqueue))
457 		wake_up_interruptible_poll(wqueue, POLLOUT | POLLWRNORM | POLLWRBAND);
458 }
459 
460 static void macvtap_sock_destruct(struct sock *sk)
461 {
462 	skb_queue_purge(&sk->sk_receive_queue);
463 }
464 
465 static int macvtap_open(struct inode *inode, struct file *file)
466 {
467 	struct net *net = current->nsproxy->net_ns;
468 	struct net_device *dev;
469 	struct macvtap_queue *q;
470 	int err = -ENODEV;
471 
472 	rtnl_lock();
473 	dev = dev_get_by_macvtap_minor(iminor(inode));
474 	if (!dev)
475 		goto out;
476 
477 	err = -ENOMEM;
478 	q = (struct macvtap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
479 					     &macvtap_proto);
480 	if (!q)
481 		goto out;
482 
483 	RCU_INIT_POINTER(q->sock.wq, &q->wq);
484 	init_waitqueue_head(&q->wq.wait);
485 	q->sock.type = SOCK_RAW;
486 	q->sock.state = SS_CONNECTED;
487 	q->sock.file = file;
488 	q->sock.ops = &macvtap_socket_ops;
489 	sock_init_data(&q->sock, &q->sk);
490 	q->sk.sk_write_space = macvtap_sock_write_space;
491 	q->sk.sk_destruct = macvtap_sock_destruct;
492 	q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
493 	q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
494 
495 	/*
496 	 * so far only KVM virtio_net uses macvtap, enable zero copy between
497 	 * guest kernel and host kernel when lower device supports zerocopy
498 	 *
499 	 * The macvlan supports zerocopy iff the lower device supports zero
500 	 * copy so we don't have to look at the lower device directly.
501 	 */
502 	if ((dev->features & NETIF_F_HIGHDMA) && (dev->features & NETIF_F_SG))
503 		sock_set_flag(&q->sk, SOCK_ZEROCOPY);
504 
505 	err = macvtap_set_queue(dev, file, q);
506 	if (err)
507 		sock_put(&q->sk);
508 
509 out:
510 	if (dev)
511 		dev_put(dev);
512 
513 	rtnl_unlock();
514 	return err;
515 }
516 
517 static int macvtap_release(struct inode *inode, struct file *file)
518 {
519 	struct macvtap_queue *q = file->private_data;
520 	macvtap_put_queue(q);
521 	return 0;
522 }
523 
524 static unsigned int macvtap_poll(struct file *file, poll_table * wait)
525 {
526 	struct macvtap_queue *q = file->private_data;
527 	unsigned int mask = POLLERR;
528 
529 	if (!q)
530 		goto out;
531 
532 	mask = 0;
533 	poll_wait(file, &q->wq.wait, wait);
534 
535 	if (!skb_queue_empty(&q->sk.sk_receive_queue))
536 		mask |= POLLIN | POLLRDNORM;
537 
538 	if (sock_writeable(&q->sk) ||
539 	    (!test_and_set_bit(SOCK_ASYNC_NOSPACE, &q->sock.flags) &&
540 	     sock_writeable(&q->sk)))
541 		mask |= POLLOUT | POLLWRNORM;
542 
543 out:
544 	return mask;
545 }
546 
547 static inline struct sk_buff *macvtap_alloc_skb(struct sock *sk, size_t prepad,
548 						size_t len, size_t linear,
549 						int noblock, int *err)
550 {
551 	struct sk_buff *skb;
552 
553 	/* Under a page?  Don't bother with paged skb. */
554 	if (prepad + len < PAGE_SIZE || !linear)
555 		linear = len;
556 
557 	skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
558 				   err, 0);
559 	if (!skb)
560 		return NULL;
561 
562 	skb_reserve(skb, prepad);
563 	skb_put(skb, linear);
564 	skb->data_len = len - linear;
565 	skb->len += len - linear;
566 
567 	return skb;
568 }
569 
570 /*
571  * macvtap_skb_from_vnet_hdr and macvtap_skb_to_vnet_hdr should
572  * be shared with the tun/tap driver.
573  */
574 static int macvtap_skb_from_vnet_hdr(struct macvtap_queue *q,
575 				     struct sk_buff *skb,
576 				     struct virtio_net_hdr *vnet_hdr)
577 {
578 	unsigned short gso_type = 0;
579 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
580 		switch (vnet_hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
581 		case VIRTIO_NET_HDR_GSO_TCPV4:
582 			gso_type = SKB_GSO_TCPV4;
583 			break;
584 		case VIRTIO_NET_HDR_GSO_TCPV6:
585 			gso_type = SKB_GSO_TCPV6;
586 			break;
587 		case VIRTIO_NET_HDR_GSO_UDP:
588 			gso_type = SKB_GSO_UDP;
589 			break;
590 		default:
591 			return -EINVAL;
592 		}
593 
594 		if (vnet_hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
595 			gso_type |= SKB_GSO_TCP_ECN;
596 
597 		if (vnet_hdr->gso_size == 0)
598 			return -EINVAL;
599 	}
600 
601 	if (vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
602 		if (!skb_partial_csum_set(skb, macvtap16_to_cpu(q, vnet_hdr->csum_start),
603 					  macvtap16_to_cpu(q, vnet_hdr->csum_offset)))
604 			return -EINVAL;
605 	}
606 
607 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
608 		skb_shinfo(skb)->gso_size = macvtap16_to_cpu(q, vnet_hdr->gso_size);
609 		skb_shinfo(skb)->gso_type = gso_type;
610 
611 		/* Header must be checked, and gso_segs computed. */
612 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
613 		skb_shinfo(skb)->gso_segs = 0;
614 	}
615 	return 0;
616 }
617 
618 static void macvtap_skb_to_vnet_hdr(struct macvtap_queue *q,
619 				    const struct sk_buff *skb,
620 				    struct virtio_net_hdr *vnet_hdr)
621 {
622 	memset(vnet_hdr, 0, sizeof(*vnet_hdr));
623 
624 	if (skb_is_gso(skb)) {
625 		struct skb_shared_info *sinfo = skb_shinfo(skb);
626 
627 		/* This is a hint as to how much should be linear. */
628 		vnet_hdr->hdr_len = cpu_to_macvtap16(q, skb_headlen(skb));
629 		vnet_hdr->gso_size = cpu_to_macvtap16(q, sinfo->gso_size);
630 		if (sinfo->gso_type & SKB_GSO_TCPV4)
631 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
632 		else if (sinfo->gso_type & SKB_GSO_TCPV6)
633 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
634 		else if (sinfo->gso_type & SKB_GSO_UDP)
635 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
636 		else
637 			BUG();
638 		if (sinfo->gso_type & SKB_GSO_TCP_ECN)
639 			vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
640 	} else
641 		vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
642 
643 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
644 		vnet_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
645 		if (skb_vlan_tag_present(skb))
646 			vnet_hdr->csum_start = cpu_to_macvtap16(q,
647 				skb_checksum_start_offset(skb) + VLAN_HLEN);
648 		else
649 			vnet_hdr->csum_start = cpu_to_macvtap16(q,
650 				skb_checksum_start_offset(skb));
651 		vnet_hdr->csum_offset = cpu_to_macvtap16(q, skb->csum_offset);
652 	} else if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
653 		vnet_hdr->flags = VIRTIO_NET_HDR_F_DATA_VALID;
654 	} /* else everything is zero */
655 }
656 
657 /* Neighbour code has some assumptions on HH_DATA_MOD alignment */
658 #define MACVTAP_RESERVE HH_DATA_OFF(ETH_HLEN)
659 
660 /* Get packet from user space buffer */
661 static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m,
662 				struct iov_iter *from, int noblock)
663 {
664 	int good_linear = SKB_MAX_HEAD(MACVTAP_RESERVE);
665 	struct sk_buff *skb;
666 	struct macvlan_dev *vlan;
667 	unsigned long total_len = iov_iter_count(from);
668 	unsigned long len = total_len;
669 	int err;
670 	struct virtio_net_hdr vnet_hdr = { 0 };
671 	int vnet_hdr_len = 0;
672 	int copylen = 0;
673 	bool zerocopy = false;
674 	size_t linear;
675 	ssize_t n;
676 
677 	if (q->flags & IFF_VNET_HDR) {
678 		vnet_hdr_len = q->vnet_hdr_sz;
679 
680 		err = -EINVAL;
681 		if (len < vnet_hdr_len)
682 			goto err;
683 		len -= vnet_hdr_len;
684 
685 		err = -EFAULT;
686 		n = copy_from_iter(&vnet_hdr, sizeof(vnet_hdr), from);
687 		if (n != sizeof(vnet_hdr))
688 			goto err;
689 		iov_iter_advance(from, vnet_hdr_len - sizeof(vnet_hdr));
690 		if ((vnet_hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
691 		     macvtap16_to_cpu(q, vnet_hdr.csum_start) +
692 		     macvtap16_to_cpu(q, vnet_hdr.csum_offset) + 2 >
693 			     macvtap16_to_cpu(q, vnet_hdr.hdr_len))
694 			vnet_hdr.hdr_len = cpu_to_macvtap16(q,
695 				 macvtap16_to_cpu(q, vnet_hdr.csum_start) +
696 				 macvtap16_to_cpu(q, vnet_hdr.csum_offset) + 2);
697 		err = -EINVAL;
698 		if (macvtap16_to_cpu(q, vnet_hdr.hdr_len) > len)
699 			goto err;
700 	}
701 
702 	err = -EINVAL;
703 	if (unlikely(len < ETH_HLEN))
704 		goto err;
705 
706 	if (m && m->msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
707 		struct iov_iter i;
708 
709 		copylen = vnet_hdr.hdr_len ?
710 			macvtap16_to_cpu(q, vnet_hdr.hdr_len) : GOODCOPY_LEN;
711 		if (copylen > good_linear)
712 			copylen = good_linear;
713 		linear = copylen;
714 		i = *from;
715 		iov_iter_advance(&i, copylen);
716 		if (iov_iter_npages(&i, INT_MAX) <= MAX_SKB_FRAGS)
717 			zerocopy = true;
718 	}
719 
720 	if (!zerocopy) {
721 		copylen = len;
722 		if (macvtap16_to_cpu(q, vnet_hdr.hdr_len) > good_linear)
723 			linear = good_linear;
724 		else
725 			linear = macvtap16_to_cpu(q, vnet_hdr.hdr_len);
726 	}
727 
728 	skb = macvtap_alloc_skb(&q->sk, MACVTAP_RESERVE, copylen,
729 				linear, noblock, &err);
730 	if (!skb)
731 		goto err;
732 
733 	if (zerocopy)
734 		err = zerocopy_sg_from_iter(skb, from);
735 	else {
736 		err = skb_copy_datagram_from_iter(skb, 0, from, len);
737 		if (!err && m && m->msg_control) {
738 			struct ubuf_info *uarg = m->msg_control;
739 			uarg->callback(uarg, false);
740 		}
741 	}
742 
743 	if (err)
744 		goto err_kfree;
745 
746 	skb_set_network_header(skb, ETH_HLEN);
747 	skb_reset_mac_header(skb);
748 	skb->protocol = eth_hdr(skb)->h_proto;
749 
750 	if (vnet_hdr_len) {
751 		err = macvtap_skb_from_vnet_hdr(q, skb, &vnet_hdr);
752 		if (err)
753 			goto err_kfree;
754 	}
755 
756 	skb_probe_transport_header(skb, ETH_HLEN);
757 
758 	rcu_read_lock();
759 	vlan = rcu_dereference(q->vlan);
760 	/* copy skb_ubuf_info for callback when skb has no error */
761 	if (zerocopy) {
762 		skb_shinfo(skb)->destructor_arg = m->msg_control;
763 		skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
764 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
765 	}
766 	if (vlan) {
767 		skb->dev = vlan->dev;
768 		dev_queue_xmit(skb);
769 	} else {
770 		kfree_skb(skb);
771 	}
772 	rcu_read_unlock();
773 
774 	return total_len;
775 
776 err_kfree:
777 	kfree_skb(skb);
778 
779 err:
780 	rcu_read_lock();
781 	vlan = rcu_dereference(q->vlan);
782 	if (vlan)
783 		this_cpu_inc(vlan->pcpu_stats->tx_dropped);
784 	rcu_read_unlock();
785 
786 	return err;
787 }
788 
789 static ssize_t macvtap_write_iter(struct kiocb *iocb, struct iov_iter *from)
790 {
791 	struct file *file = iocb->ki_filp;
792 	struct macvtap_queue *q = file->private_data;
793 
794 	return macvtap_get_user(q, NULL, from, file->f_flags & O_NONBLOCK);
795 }
796 
797 /* Put packet to the user space buffer */
798 static ssize_t macvtap_put_user(struct macvtap_queue *q,
799 				const struct sk_buff *skb,
800 				struct iov_iter *iter)
801 {
802 	int ret;
803 	int vnet_hdr_len = 0;
804 	int vlan_offset = 0;
805 	int total;
806 
807 	if (q->flags & IFF_VNET_HDR) {
808 		struct virtio_net_hdr vnet_hdr;
809 		vnet_hdr_len = q->vnet_hdr_sz;
810 		if (iov_iter_count(iter) < vnet_hdr_len)
811 			return -EINVAL;
812 
813 		macvtap_skb_to_vnet_hdr(q, skb, &vnet_hdr);
814 
815 		if (copy_to_iter(&vnet_hdr, sizeof(vnet_hdr), iter) !=
816 		    sizeof(vnet_hdr))
817 			return -EFAULT;
818 
819 		iov_iter_advance(iter, vnet_hdr_len - sizeof(vnet_hdr));
820 	}
821 	total = vnet_hdr_len;
822 	total += skb->len;
823 
824 	if (skb_vlan_tag_present(skb)) {
825 		struct {
826 			__be16 h_vlan_proto;
827 			__be16 h_vlan_TCI;
828 		} veth;
829 		veth.h_vlan_proto = skb->vlan_proto;
830 		veth.h_vlan_TCI = htons(skb_vlan_tag_get(skb));
831 
832 		vlan_offset = offsetof(struct vlan_ethhdr, h_vlan_proto);
833 		total += VLAN_HLEN;
834 
835 		ret = skb_copy_datagram_iter(skb, 0, iter, vlan_offset);
836 		if (ret || !iov_iter_count(iter))
837 			goto done;
838 
839 		ret = copy_to_iter(&veth, sizeof(veth), iter);
840 		if (ret != sizeof(veth) || !iov_iter_count(iter))
841 			goto done;
842 	}
843 
844 	ret = skb_copy_datagram_iter(skb, vlan_offset, iter,
845 				     skb->len - vlan_offset);
846 
847 done:
848 	return ret ? ret : total;
849 }
850 
851 static ssize_t macvtap_do_read(struct macvtap_queue *q,
852 			       struct iov_iter *to,
853 			       int noblock)
854 {
855 	DEFINE_WAIT(wait);
856 	struct sk_buff *skb;
857 	ssize_t ret = 0;
858 
859 	if (!iov_iter_count(to))
860 		return 0;
861 
862 	while (1) {
863 		if (!noblock)
864 			prepare_to_wait(sk_sleep(&q->sk), &wait,
865 					TASK_INTERRUPTIBLE);
866 
867 		/* Read frames from the queue */
868 		skb = skb_dequeue(&q->sk.sk_receive_queue);
869 		if (skb)
870 			break;
871 		if (noblock) {
872 			ret = -EAGAIN;
873 			break;
874 		}
875 		if (signal_pending(current)) {
876 			ret = -ERESTARTSYS;
877 			break;
878 		}
879 		/* Nothing to read, let's sleep */
880 		schedule();
881 	}
882 	if (skb) {
883 		ret = macvtap_put_user(q, skb, to);
884 		if (unlikely(ret < 0))
885 			kfree_skb(skb);
886 		else
887 			consume_skb(skb);
888 	}
889 	if (!noblock)
890 		finish_wait(sk_sleep(&q->sk), &wait);
891 	return ret;
892 }
893 
894 static ssize_t macvtap_read_iter(struct kiocb *iocb, struct iov_iter *to)
895 {
896 	struct file *file = iocb->ki_filp;
897 	struct macvtap_queue *q = file->private_data;
898 	ssize_t len = iov_iter_count(to), ret;
899 
900 	ret = macvtap_do_read(q, to, file->f_flags & O_NONBLOCK);
901 	ret = min_t(ssize_t, ret, len);
902 	if (ret > 0)
903 		iocb->ki_pos = ret;
904 	return ret;
905 }
906 
907 static struct macvlan_dev *macvtap_get_vlan(struct macvtap_queue *q)
908 {
909 	struct macvlan_dev *vlan;
910 
911 	ASSERT_RTNL();
912 	vlan = rtnl_dereference(q->vlan);
913 	if (vlan)
914 		dev_hold(vlan->dev);
915 
916 	return vlan;
917 }
918 
919 static void macvtap_put_vlan(struct macvlan_dev *vlan)
920 {
921 	dev_put(vlan->dev);
922 }
923 
924 static int macvtap_ioctl_set_queue(struct file *file, unsigned int flags)
925 {
926 	struct macvtap_queue *q = file->private_data;
927 	struct macvlan_dev *vlan;
928 	int ret;
929 
930 	vlan = macvtap_get_vlan(q);
931 	if (!vlan)
932 		return -EINVAL;
933 
934 	if (flags & IFF_ATTACH_QUEUE)
935 		ret = macvtap_enable_queue(vlan->dev, file, q);
936 	else if (flags & IFF_DETACH_QUEUE)
937 		ret = macvtap_disable_queue(q);
938 	else
939 		ret = -EINVAL;
940 
941 	macvtap_put_vlan(vlan);
942 	return ret;
943 }
944 
945 static int set_offload(struct macvtap_queue *q, unsigned long arg)
946 {
947 	struct macvlan_dev *vlan;
948 	netdev_features_t features;
949 	netdev_features_t feature_mask = 0;
950 
951 	vlan = rtnl_dereference(q->vlan);
952 	if (!vlan)
953 		return -ENOLINK;
954 
955 	features = vlan->dev->features;
956 
957 	if (arg & TUN_F_CSUM) {
958 		feature_mask = NETIF_F_HW_CSUM;
959 
960 		if (arg & (TUN_F_TSO4 | TUN_F_TSO6)) {
961 			if (arg & TUN_F_TSO_ECN)
962 				feature_mask |= NETIF_F_TSO_ECN;
963 			if (arg & TUN_F_TSO4)
964 				feature_mask |= NETIF_F_TSO;
965 			if (arg & TUN_F_TSO6)
966 				feature_mask |= NETIF_F_TSO6;
967 		}
968 
969 		if (arg & TUN_F_UFO)
970 			feature_mask |= NETIF_F_UFO;
971 	}
972 
973 	/* tun/tap driver inverts the usage for TSO offloads, where
974 	 * setting the TSO bit means that the userspace wants to
975 	 * accept TSO frames and turning it off means that user space
976 	 * does not support TSO.
977 	 * For macvtap, we have to invert it to mean the same thing.
978 	 * When user space turns off TSO, we turn off GSO/LRO so that
979 	 * user-space will not receive TSO frames.
980 	 */
981 	if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6 | NETIF_F_UFO))
982 		features |= RX_OFFLOADS;
983 	else
984 		features &= ~RX_OFFLOADS;
985 
986 	/* tap_features are the same as features on tun/tap and
987 	 * reflect user expectations.
988 	 */
989 	vlan->tap_features = feature_mask;
990 	vlan->set_features = features;
991 	netdev_update_features(vlan->dev);
992 
993 	return 0;
994 }
995 
996 /*
997  * provide compatibility with generic tun/tap interface
998  */
999 static long macvtap_ioctl(struct file *file, unsigned int cmd,
1000 			  unsigned long arg)
1001 {
1002 	struct macvtap_queue *q = file->private_data;
1003 	struct macvlan_dev *vlan;
1004 	void __user *argp = (void __user *)arg;
1005 	struct ifreq __user *ifr = argp;
1006 	unsigned int __user *up = argp;
1007 	unsigned short u;
1008 	int __user *sp = argp;
1009 	int s;
1010 	int ret;
1011 
1012 	switch (cmd) {
1013 	case TUNSETIFF:
1014 		/* ignore the name, just look at flags */
1015 		if (get_user(u, &ifr->ifr_flags))
1016 			return -EFAULT;
1017 
1018 		ret = 0;
1019 		if ((u & ~MACVTAP_FEATURES) != (IFF_NO_PI | IFF_TAP))
1020 			ret = -EINVAL;
1021 		else
1022 			q->flags = (q->flags & ~MACVTAP_FEATURES) | u;
1023 
1024 		return ret;
1025 
1026 	case TUNGETIFF:
1027 		rtnl_lock();
1028 		vlan = macvtap_get_vlan(q);
1029 		if (!vlan) {
1030 			rtnl_unlock();
1031 			return -ENOLINK;
1032 		}
1033 
1034 		ret = 0;
1035 		u = q->flags;
1036 		if (copy_to_user(&ifr->ifr_name, vlan->dev->name, IFNAMSIZ) ||
1037 		    put_user(u, &ifr->ifr_flags))
1038 			ret = -EFAULT;
1039 		macvtap_put_vlan(vlan);
1040 		rtnl_unlock();
1041 		return ret;
1042 
1043 	case TUNSETQUEUE:
1044 		if (get_user(u, &ifr->ifr_flags))
1045 			return -EFAULT;
1046 		rtnl_lock();
1047 		ret = macvtap_ioctl_set_queue(file, u);
1048 		rtnl_unlock();
1049 		return ret;
1050 
1051 	case TUNGETFEATURES:
1052 		if (put_user(IFF_TAP | IFF_NO_PI | MACVTAP_FEATURES, up))
1053 			return -EFAULT;
1054 		return 0;
1055 
1056 	case TUNSETSNDBUF:
1057 		if (get_user(u, up))
1058 			return -EFAULT;
1059 
1060 		q->sk.sk_sndbuf = u;
1061 		return 0;
1062 
1063 	case TUNGETVNETHDRSZ:
1064 		s = q->vnet_hdr_sz;
1065 		if (put_user(s, sp))
1066 			return -EFAULT;
1067 		return 0;
1068 
1069 	case TUNSETVNETHDRSZ:
1070 		if (get_user(s, sp))
1071 			return -EFAULT;
1072 		if (s < (int)sizeof(struct virtio_net_hdr))
1073 			return -EINVAL;
1074 
1075 		q->vnet_hdr_sz = s;
1076 		return 0;
1077 
1078 	case TUNGETVNETLE:
1079 		s = !!(q->flags & MACVTAP_VNET_LE);
1080 		if (put_user(s, sp))
1081 			return -EFAULT;
1082 		return 0;
1083 
1084 	case TUNSETVNETLE:
1085 		if (get_user(s, sp))
1086 			return -EFAULT;
1087 		if (s)
1088 			q->flags |= MACVTAP_VNET_LE;
1089 		else
1090 			q->flags &= ~MACVTAP_VNET_LE;
1091 		return 0;
1092 
1093 	case TUNSETOFFLOAD:
1094 		/* let the user check for future flags */
1095 		if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1096 			    TUN_F_TSO_ECN | TUN_F_UFO))
1097 			return -EINVAL;
1098 
1099 		rtnl_lock();
1100 		ret = set_offload(q, arg);
1101 		rtnl_unlock();
1102 		return ret;
1103 
1104 	default:
1105 		return -EINVAL;
1106 	}
1107 }
1108 
1109 #ifdef CONFIG_COMPAT
1110 static long macvtap_compat_ioctl(struct file *file, unsigned int cmd,
1111 				 unsigned long arg)
1112 {
1113 	return macvtap_ioctl(file, cmd, (unsigned long)compat_ptr(arg));
1114 }
1115 #endif
1116 
1117 static const struct file_operations macvtap_fops = {
1118 	.owner		= THIS_MODULE,
1119 	.open		= macvtap_open,
1120 	.release	= macvtap_release,
1121 	.read		= new_sync_read,
1122 	.write		= new_sync_write,
1123 	.read_iter	= macvtap_read_iter,
1124 	.write_iter	= macvtap_write_iter,
1125 	.poll		= macvtap_poll,
1126 	.llseek		= no_llseek,
1127 	.unlocked_ioctl	= macvtap_ioctl,
1128 #ifdef CONFIG_COMPAT
1129 	.compat_ioctl	= macvtap_compat_ioctl,
1130 #endif
1131 };
1132 
1133 static int macvtap_sendmsg(struct kiocb *iocb, struct socket *sock,
1134 			   struct msghdr *m, size_t total_len)
1135 {
1136 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1137 	return macvtap_get_user(q, m, &m->msg_iter, m->msg_flags & MSG_DONTWAIT);
1138 }
1139 
1140 static int macvtap_recvmsg(struct kiocb *iocb, struct socket *sock,
1141 			   struct msghdr *m, size_t total_len,
1142 			   int flags)
1143 {
1144 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1145 	int ret;
1146 	if (flags & ~(MSG_DONTWAIT|MSG_TRUNC))
1147 		return -EINVAL;
1148 	ret = macvtap_do_read(q, &m->msg_iter, flags & MSG_DONTWAIT);
1149 	if (ret > total_len) {
1150 		m->msg_flags |= MSG_TRUNC;
1151 		ret = flags & MSG_TRUNC ? ret : total_len;
1152 	}
1153 	return ret;
1154 }
1155 
1156 /* Ops structure to mimic raw sockets with tun */
1157 static const struct proto_ops macvtap_socket_ops = {
1158 	.sendmsg = macvtap_sendmsg,
1159 	.recvmsg = macvtap_recvmsg,
1160 };
1161 
1162 /* Get an underlying socket object from tun file.  Returns error unless file is
1163  * attached to a device.  The returned object works like a packet socket, it
1164  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1165  * holding a reference to the file for as long as the socket is in use. */
1166 struct socket *macvtap_get_socket(struct file *file)
1167 {
1168 	struct macvtap_queue *q;
1169 	if (file->f_op != &macvtap_fops)
1170 		return ERR_PTR(-EINVAL);
1171 	q = file->private_data;
1172 	if (!q)
1173 		return ERR_PTR(-EBADFD);
1174 	return &q->sock;
1175 }
1176 EXPORT_SYMBOL_GPL(macvtap_get_socket);
1177 
1178 static int macvtap_device_event(struct notifier_block *unused,
1179 				unsigned long event, void *ptr)
1180 {
1181 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1182 	struct macvlan_dev *vlan;
1183 	struct device *classdev;
1184 	dev_t devt;
1185 	int err;
1186 
1187 	if (dev->rtnl_link_ops != &macvtap_link_ops)
1188 		return NOTIFY_DONE;
1189 
1190 	vlan = netdev_priv(dev);
1191 
1192 	switch (event) {
1193 	case NETDEV_REGISTER:
1194 		/* Create the device node here after the network device has
1195 		 * been registered but before register_netdevice has
1196 		 * finished running.
1197 		 */
1198 		err = macvtap_get_minor(vlan);
1199 		if (err)
1200 			return notifier_from_errno(err);
1201 
1202 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1203 		classdev = device_create(macvtap_class, &dev->dev, devt,
1204 					 dev, "tap%d", dev->ifindex);
1205 		if (IS_ERR(classdev)) {
1206 			macvtap_free_minor(vlan);
1207 			return notifier_from_errno(PTR_ERR(classdev));
1208 		}
1209 		break;
1210 	case NETDEV_UNREGISTER:
1211 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1212 		device_destroy(macvtap_class, devt);
1213 		macvtap_free_minor(vlan);
1214 		break;
1215 	}
1216 
1217 	return NOTIFY_DONE;
1218 }
1219 
1220 static struct notifier_block macvtap_notifier_block __read_mostly = {
1221 	.notifier_call	= macvtap_device_event,
1222 };
1223 
1224 static int macvtap_init(void)
1225 {
1226 	int err;
1227 
1228 	err = alloc_chrdev_region(&macvtap_major, 0,
1229 				MACVTAP_NUM_DEVS, "macvtap");
1230 	if (err)
1231 		goto out1;
1232 
1233 	cdev_init(&macvtap_cdev, &macvtap_fops);
1234 	err = cdev_add(&macvtap_cdev, macvtap_major, MACVTAP_NUM_DEVS);
1235 	if (err)
1236 		goto out2;
1237 
1238 	macvtap_class = class_create(THIS_MODULE, "macvtap");
1239 	if (IS_ERR(macvtap_class)) {
1240 		err = PTR_ERR(macvtap_class);
1241 		goto out3;
1242 	}
1243 
1244 	err = register_netdevice_notifier(&macvtap_notifier_block);
1245 	if (err)
1246 		goto out4;
1247 
1248 	err = macvlan_link_register(&macvtap_link_ops);
1249 	if (err)
1250 		goto out5;
1251 
1252 	return 0;
1253 
1254 out5:
1255 	unregister_netdevice_notifier(&macvtap_notifier_block);
1256 out4:
1257 	class_unregister(macvtap_class);
1258 out3:
1259 	cdev_del(&macvtap_cdev);
1260 out2:
1261 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1262 out1:
1263 	return err;
1264 }
1265 module_init(macvtap_init);
1266 
1267 static void macvtap_exit(void)
1268 {
1269 	rtnl_link_unregister(&macvtap_link_ops);
1270 	unregister_netdevice_notifier(&macvtap_notifier_block);
1271 	class_unregister(macvtap_class);
1272 	cdev_del(&macvtap_cdev);
1273 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1274 }
1275 module_exit(macvtap_exit);
1276 
1277 MODULE_ALIAS_RTNL_LINK("macvtap");
1278 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1279 MODULE_LICENSE("GPL");
1280