xref: /openbmc/linux/drivers/net/macvtap.c (revision c819e2cf)
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/ipv6.h>
21 #include <net/net_namespace.h>
22 #include <net/rtnetlink.h>
23 #include <net/sock.h>
24 #include <linux/virtio_net.h>
25 
26 /*
27  * A macvtap queue is the central object of this driver, it connects
28  * an open character device to a macvlan interface. There can be
29  * multiple queues on one interface, which map back to queues
30  * implemented in hardware on the underlying device.
31  *
32  * macvtap_proto is used to allocate queues through the sock allocation
33  * mechanism.
34  *
35  */
36 struct macvtap_queue {
37 	struct sock sk;
38 	struct socket sock;
39 	struct socket_wq wq;
40 	int vnet_hdr_sz;
41 	struct macvlan_dev __rcu *vlan;
42 	struct file *file;
43 	unsigned int flags;
44 	u16 queue_index;
45 	bool enabled;
46 	struct list_head next;
47 };
48 
49 #define MACVTAP_FEATURES (IFF_VNET_HDR | IFF_MULTI_QUEUE)
50 
51 #define MACVTAP_VNET_LE 0x80000000
52 
53 static inline u16 macvtap16_to_cpu(struct macvtap_queue *q, __virtio16 val)
54 {
55 	return __virtio16_to_cpu(q->flags & MACVTAP_VNET_LE, val);
56 }
57 
58 static inline __virtio16 cpu_to_macvtap16(struct macvtap_queue *q, u16 val)
59 {
60 	return __cpu_to_virtio16(q->flags & MACVTAP_VNET_LE, val);
61 }
62 
63 static struct proto macvtap_proto = {
64 	.name = "macvtap",
65 	.owner = THIS_MODULE,
66 	.obj_size = sizeof (struct macvtap_queue),
67 };
68 
69 /*
70  * Variables for dealing with macvtaps device numbers.
71  */
72 static dev_t macvtap_major;
73 #define MACVTAP_NUM_DEVS (1U << MINORBITS)
74 static DEFINE_MUTEX(minor_lock);
75 static DEFINE_IDR(minor_idr);
76 
77 #define GOODCOPY_LEN 128
78 static struct class *macvtap_class;
79 static struct cdev macvtap_cdev;
80 
81 static const struct proto_ops macvtap_socket_ops;
82 
83 #define TUN_OFFLOADS (NETIF_F_HW_CSUM | NETIF_F_TSO_ECN | NETIF_F_TSO | \
84 		      NETIF_F_TSO6)
85 #define RX_OFFLOADS (NETIF_F_GRO | NETIF_F_LRO)
86 #define TAP_FEATURES (NETIF_F_GSO | NETIF_F_SG)
87 
88 static struct macvlan_dev *macvtap_get_vlan_rcu(const struct net_device *dev)
89 {
90 	return rcu_dereference(dev->rx_handler_data);
91 }
92 
93 /*
94  * RCU usage:
95  * The macvtap_queue and the macvlan_dev are loosely coupled, the
96  * pointers from one to the other can only be read while rcu_read_lock
97  * or rtnl is held.
98  *
99  * Both the file and the macvlan_dev hold a reference on the macvtap_queue
100  * through sock_hold(&q->sk). When the macvlan_dev goes away first,
101  * q->vlan becomes inaccessible. When the files gets closed,
102  * macvtap_get_queue() fails.
103  *
104  * There may still be references to the struct sock inside of the
105  * queue from outbound SKBs, but these never reference back to the
106  * file or the dev. The data structure is freed through __sk_free
107  * when both our references and any pending SKBs are gone.
108  */
109 
110 static int macvtap_enable_queue(struct net_device *dev, struct file *file,
111 				struct macvtap_queue *q)
112 {
113 	struct macvlan_dev *vlan = netdev_priv(dev);
114 	int err = -EINVAL;
115 
116 	ASSERT_RTNL();
117 
118 	if (q->enabled)
119 		goto out;
120 
121 	err = 0;
122 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
123 	q->queue_index = vlan->numvtaps;
124 	q->enabled = true;
125 
126 	vlan->numvtaps++;
127 out:
128 	return err;
129 }
130 
131 /* Requires RTNL */
132 static int macvtap_set_queue(struct net_device *dev, struct file *file,
133 			     struct macvtap_queue *q)
134 {
135 	struct macvlan_dev *vlan = netdev_priv(dev);
136 
137 	if (vlan->numqueues == MAX_MACVTAP_QUEUES)
138 		return -EBUSY;
139 
140 	rcu_assign_pointer(q->vlan, vlan);
141 	rcu_assign_pointer(vlan->taps[vlan->numvtaps], q);
142 	sock_hold(&q->sk);
143 
144 	q->file = file;
145 	q->queue_index = vlan->numvtaps;
146 	q->enabled = true;
147 	file->private_data = q;
148 	list_add_tail(&q->next, &vlan->queue_list);
149 
150 	vlan->numvtaps++;
151 	vlan->numqueues++;
152 
153 	return 0;
154 }
155 
156 static int macvtap_disable_queue(struct macvtap_queue *q)
157 {
158 	struct macvlan_dev *vlan;
159 	struct macvtap_queue *nq;
160 
161 	ASSERT_RTNL();
162 	if (!q->enabled)
163 		return -EINVAL;
164 
165 	vlan = rtnl_dereference(q->vlan);
166 
167 	if (vlan) {
168 		int index = q->queue_index;
169 		BUG_ON(index >= vlan->numvtaps);
170 		nq = rtnl_dereference(vlan->taps[vlan->numvtaps - 1]);
171 		nq->queue_index = index;
172 
173 		rcu_assign_pointer(vlan->taps[index], nq);
174 		RCU_INIT_POINTER(vlan->taps[vlan->numvtaps - 1], NULL);
175 		q->enabled = false;
176 
177 		vlan->numvtaps--;
178 	}
179 
180 	return 0;
181 }
182 
183 /*
184  * The file owning the queue got closed, give up both
185  * the reference that the files holds as well as the
186  * one from the macvlan_dev if that still exists.
187  *
188  * Using the spinlock makes sure that we don't get
189  * to the queue again after destroying it.
190  */
191 static void macvtap_put_queue(struct macvtap_queue *q)
192 {
193 	struct macvlan_dev *vlan;
194 
195 	rtnl_lock();
196 	vlan = rtnl_dereference(q->vlan);
197 
198 	if (vlan) {
199 		if (q->enabled)
200 			BUG_ON(macvtap_disable_queue(q));
201 
202 		vlan->numqueues--;
203 		RCU_INIT_POINTER(q->vlan, NULL);
204 		sock_put(&q->sk);
205 		list_del_init(&q->next);
206 	}
207 
208 	rtnl_unlock();
209 
210 	synchronize_rcu();
211 	sock_put(&q->sk);
212 }
213 
214 /*
215  * Select a queue based on the rxq of the device on which this packet
216  * arrived. If the incoming device is not mq, calculate a flow hash
217  * to select a queue. If all fails, find the first available queue.
218  * Cache vlan->numvtaps since it can become zero during the execution
219  * of this function.
220  */
221 static struct macvtap_queue *macvtap_get_queue(struct net_device *dev,
222 					       struct sk_buff *skb)
223 {
224 	struct macvlan_dev *vlan = netdev_priv(dev);
225 	struct macvtap_queue *tap = NULL;
226 	/* Access to taps array is protected by rcu, but access to numvtaps
227 	 * isn't. Below we use it to lookup a queue, but treat it as a hint
228 	 * and validate that the result isn't NULL - in case we are
229 	 * racing against queue removal.
230 	 */
231 	int numvtaps = ACCESS_ONCE(vlan->numvtaps);
232 	__u32 rxq;
233 
234 	if (!numvtaps)
235 		goto out;
236 
237 	/* Check if we can use flow to select a queue */
238 	rxq = skb_get_hash(skb);
239 	if (rxq) {
240 		tap = rcu_dereference(vlan->taps[rxq % numvtaps]);
241 		goto out;
242 	}
243 
244 	if (likely(skb_rx_queue_recorded(skb))) {
245 		rxq = skb_get_rx_queue(skb);
246 
247 		while (unlikely(rxq >= numvtaps))
248 			rxq -= numvtaps;
249 
250 		tap = rcu_dereference(vlan->taps[rxq]);
251 		goto out;
252 	}
253 
254 	tap = rcu_dereference(vlan->taps[0]);
255 out:
256 	return tap;
257 }
258 
259 /*
260  * The net_device is going away, give up the reference
261  * that it holds on all queues and safely set the pointer
262  * from the queues to NULL.
263  */
264 static void macvtap_del_queues(struct net_device *dev)
265 {
266 	struct macvlan_dev *vlan = netdev_priv(dev);
267 	struct macvtap_queue *q, *tmp, *qlist[MAX_MACVTAP_QUEUES];
268 	int i, j = 0;
269 
270 	ASSERT_RTNL();
271 	list_for_each_entry_safe(q, tmp, &vlan->queue_list, next) {
272 		list_del_init(&q->next);
273 		qlist[j++] = q;
274 		RCU_INIT_POINTER(q->vlan, NULL);
275 		if (q->enabled)
276 			vlan->numvtaps--;
277 		vlan->numqueues--;
278 	}
279 	for (i = 0; i < vlan->numvtaps; i++)
280 		RCU_INIT_POINTER(vlan->taps[i], NULL);
281 	BUG_ON(vlan->numvtaps);
282 	BUG_ON(vlan->numqueues);
283 	/* guarantee that any future macvtap_set_queue will fail */
284 	vlan->numvtaps = MAX_MACVTAP_QUEUES;
285 
286 	for (--j; j >= 0; j--)
287 		sock_put(&qlist[j]->sk);
288 }
289 
290 static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
291 {
292 	struct sk_buff *skb = *pskb;
293 	struct net_device *dev = skb->dev;
294 	struct macvlan_dev *vlan;
295 	struct macvtap_queue *q;
296 	netdev_features_t features = TAP_FEATURES;
297 
298 	vlan = macvtap_get_vlan_rcu(dev);
299 	if (!vlan)
300 		return RX_HANDLER_PASS;
301 
302 	q = macvtap_get_queue(dev, skb);
303 	if (!q)
304 		return RX_HANDLER_PASS;
305 
306 	if (skb_queue_len(&q->sk.sk_receive_queue) >= dev->tx_queue_len)
307 		goto drop;
308 
309 	skb_push(skb, ETH_HLEN);
310 
311 	/* Apply the forward feature mask so that we perform segmentation
312 	 * according to users wishes.  This only works if VNET_HDR is
313 	 * enabled.
314 	 */
315 	if (q->flags & IFF_VNET_HDR)
316 		features |= vlan->tap_features;
317 	if (netif_needs_gso(dev, skb, features)) {
318 		struct sk_buff *segs = __skb_gso_segment(skb, features, false);
319 
320 		if (IS_ERR(segs))
321 			goto drop;
322 
323 		if (!segs) {
324 			skb_queue_tail(&q->sk.sk_receive_queue, skb);
325 			goto wake_up;
326 		}
327 
328 		kfree_skb(skb);
329 		while (segs) {
330 			struct sk_buff *nskb = segs->next;
331 
332 			segs->next = NULL;
333 			skb_queue_tail(&q->sk.sk_receive_queue, segs);
334 			segs = nskb;
335 		}
336 	} else {
337 		/* If we receive a partial checksum and the tap side
338 		 * doesn't support checksum offload, compute the checksum.
339 		 * Note: it doesn't matter which checksum feature to
340 		 *        check, we either support them all or none.
341 		 */
342 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
343 		    !(features & NETIF_F_ALL_CSUM) &&
344 		    skb_checksum_help(skb))
345 			goto drop;
346 		skb_queue_tail(&q->sk.sk_receive_queue, skb);
347 	}
348 
349 wake_up:
350 	wake_up_interruptible_poll(sk_sleep(&q->sk), POLLIN | POLLRDNORM | POLLRDBAND);
351 	return RX_HANDLER_CONSUMED;
352 
353 drop:
354 	/* Count errors/drops only here, thus don't care about args. */
355 	macvlan_count_rx(vlan, 0, 0, 0);
356 	kfree_skb(skb);
357 	return RX_HANDLER_CONSUMED;
358 }
359 
360 static int macvtap_get_minor(struct macvlan_dev *vlan)
361 {
362 	int retval = -ENOMEM;
363 
364 	mutex_lock(&minor_lock);
365 	retval = idr_alloc(&minor_idr, vlan, 1, MACVTAP_NUM_DEVS, GFP_KERNEL);
366 	if (retval >= 0) {
367 		vlan->minor = retval;
368 	} else if (retval == -ENOSPC) {
369 		printk(KERN_ERR "too many macvtap devices\n");
370 		retval = -EINVAL;
371 	}
372 	mutex_unlock(&minor_lock);
373 	return retval < 0 ? retval : 0;
374 }
375 
376 static void macvtap_free_minor(struct macvlan_dev *vlan)
377 {
378 	mutex_lock(&minor_lock);
379 	if (vlan->minor) {
380 		idr_remove(&minor_idr, vlan->minor);
381 		vlan->minor = 0;
382 	}
383 	mutex_unlock(&minor_lock);
384 }
385 
386 static struct net_device *dev_get_by_macvtap_minor(int minor)
387 {
388 	struct net_device *dev = NULL;
389 	struct macvlan_dev *vlan;
390 
391 	mutex_lock(&minor_lock);
392 	vlan = idr_find(&minor_idr, minor);
393 	if (vlan) {
394 		dev = vlan->dev;
395 		dev_hold(dev);
396 	}
397 	mutex_unlock(&minor_lock);
398 	return dev;
399 }
400 
401 static int macvtap_newlink(struct net *src_net,
402 			   struct net_device *dev,
403 			   struct nlattr *tb[],
404 			   struct nlattr *data[])
405 {
406 	struct macvlan_dev *vlan = netdev_priv(dev);
407 	int err;
408 
409 	INIT_LIST_HEAD(&vlan->queue_list);
410 
411 	/* Since macvlan supports all offloads by default, make
412 	 * tap support all offloads also.
413 	 */
414 	vlan->tap_features = TUN_OFFLOADS;
415 
416 	err = netdev_rx_handler_register(dev, macvtap_handle_frame, vlan);
417 	if (err)
418 		return err;
419 
420 	/* Don't put anything that may fail after macvlan_common_newlink
421 	 * because we can't undo what it does.
422 	 */
423 	return macvlan_common_newlink(src_net, dev, tb, data);
424 }
425 
426 static void macvtap_dellink(struct net_device *dev,
427 			    struct list_head *head)
428 {
429 	netdev_rx_handler_unregister(dev);
430 	macvtap_del_queues(dev);
431 	macvlan_dellink(dev, head);
432 }
433 
434 static void macvtap_setup(struct net_device *dev)
435 {
436 	macvlan_common_setup(dev);
437 	dev->tx_queue_len = TUN_READQ_SIZE;
438 }
439 
440 static struct rtnl_link_ops macvtap_link_ops __read_mostly = {
441 	.kind		= "macvtap",
442 	.setup		= macvtap_setup,
443 	.newlink	= macvtap_newlink,
444 	.dellink	= macvtap_dellink,
445 };
446 
447 
448 static void macvtap_sock_write_space(struct sock *sk)
449 {
450 	wait_queue_head_t *wqueue;
451 
452 	if (!sock_writeable(sk) ||
453 	    !test_and_clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags))
454 		return;
455 
456 	wqueue = sk_sleep(sk);
457 	if (wqueue && waitqueue_active(wqueue))
458 		wake_up_interruptible_poll(wqueue, POLLOUT | POLLWRNORM | POLLWRBAND);
459 }
460 
461 static void macvtap_sock_destruct(struct sock *sk)
462 {
463 	skb_queue_purge(&sk->sk_receive_queue);
464 }
465 
466 static int macvtap_open(struct inode *inode, struct file *file)
467 {
468 	struct net *net = current->nsproxy->net_ns;
469 	struct net_device *dev;
470 	struct macvtap_queue *q;
471 	int err = -ENODEV;
472 
473 	rtnl_lock();
474 	dev = dev_get_by_macvtap_minor(iminor(inode));
475 	if (!dev)
476 		goto out;
477 
478 	err = -ENOMEM;
479 	q = (struct macvtap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
480 					     &macvtap_proto);
481 	if (!q)
482 		goto out;
483 
484 	RCU_INIT_POINTER(q->sock.wq, &q->wq);
485 	init_waitqueue_head(&q->wq.wait);
486 	q->sock.type = SOCK_RAW;
487 	q->sock.state = SS_CONNECTED;
488 	q->sock.file = file;
489 	q->sock.ops = &macvtap_socket_ops;
490 	sock_init_data(&q->sock, &q->sk);
491 	q->sk.sk_write_space = macvtap_sock_write_space;
492 	q->sk.sk_destruct = macvtap_sock_destruct;
493 	q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
494 	q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
495 
496 	/*
497 	 * so far only KVM virtio_net uses macvtap, enable zero copy between
498 	 * guest kernel and host kernel when lower device supports zerocopy
499 	 *
500 	 * The macvlan supports zerocopy iff the lower device supports zero
501 	 * copy so we don't have to look at the lower device directly.
502 	 */
503 	if ((dev->features & NETIF_F_HIGHDMA) && (dev->features & NETIF_F_SG))
504 		sock_set_flag(&q->sk, SOCK_ZEROCOPY);
505 
506 	err = macvtap_set_queue(dev, file, q);
507 	if (err)
508 		sock_put(&q->sk);
509 
510 out:
511 	if (dev)
512 		dev_put(dev);
513 
514 	rtnl_unlock();
515 	return err;
516 }
517 
518 static int macvtap_release(struct inode *inode, struct file *file)
519 {
520 	struct macvtap_queue *q = file->private_data;
521 	macvtap_put_queue(q);
522 	return 0;
523 }
524 
525 static unsigned int macvtap_poll(struct file *file, poll_table * wait)
526 {
527 	struct macvtap_queue *q = file->private_data;
528 	unsigned int mask = POLLERR;
529 
530 	if (!q)
531 		goto out;
532 
533 	mask = 0;
534 	poll_wait(file, &q->wq.wait, wait);
535 
536 	if (!skb_queue_empty(&q->sk.sk_receive_queue))
537 		mask |= POLLIN | POLLRDNORM;
538 
539 	if (sock_writeable(&q->sk) ||
540 	    (!test_and_set_bit(SOCK_ASYNC_NOSPACE, &q->sock.flags) &&
541 	     sock_writeable(&q->sk)))
542 		mask |= POLLOUT | POLLWRNORM;
543 
544 out:
545 	return mask;
546 }
547 
548 static inline struct sk_buff *macvtap_alloc_skb(struct sock *sk, size_t prepad,
549 						size_t len, size_t linear,
550 						int noblock, int *err)
551 {
552 	struct sk_buff *skb;
553 
554 	/* Under a page?  Don't bother with paged skb. */
555 	if (prepad + len < PAGE_SIZE || !linear)
556 		linear = len;
557 
558 	skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
559 				   err, 0);
560 	if (!skb)
561 		return NULL;
562 
563 	skb_reserve(skb, prepad);
564 	skb_put(skb, linear);
565 	skb->data_len = len - linear;
566 	skb->len += len - linear;
567 
568 	return skb;
569 }
570 
571 /*
572  * macvtap_skb_from_vnet_hdr and macvtap_skb_to_vnet_hdr should
573  * be shared with the tun/tap driver.
574  */
575 static int macvtap_skb_from_vnet_hdr(struct macvtap_queue *q,
576 				     struct sk_buff *skb,
577 				     struct virtio_net_hdr *vnet_hdr)
578 {
579 	unsigned short gso_type = 0;
580 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
581 		switch (vnet_hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
582 		case VIRTIO_NET_HDR_GSO_TCPV4:
583 			gso_type = SKB_GSO_TCPV4;
584 			break;
585 		case VIRTIO_NET_HDR_GSO_TCPV6:
586 			gso_type = SKB_GSO_TCPV6;
587 			break;
588 		case VIRTIO_NET_HDR_GSO_UDP:
589 			pr_warn_once("macvtap: %s: using disabled UFO feature; please fix this program\n",
590 				     current->comm);
591 			gso_type = SKB_GSO_UDP;
592 			if (skb->protocol == htons(ETH_P_IPV6))
593 				ipv6_proxy_select_ident(skb);
594 			break;
595 		default:
596 			return -EINVAL;
597 		}
598 
599 		if (vnet_hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
600 			gso_type |= SKB_GSO_TCP_ECN;
601 
602 		if (vnet_hdr->gso_size == 0)
603 			return -EINVAL;
604 	}
605 
606 	if (vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
607 		if (!skb_partial_csum_set(skb, macvtap16_to_cpu(q, vnet_hdr->csum_start),
608 					  macvtap16_to_cpu(q, vnet_hdr->csum_offset)))
609 			return -EINVAL;
610 	}
611 
612 	if (vnet_hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
613 		skb_shinfo(skb)->gso_size = macvtap16_to_cpu(q, vnet_hdr->gso_size);
614 		skb_shinfo(skb)->gso_type = gso_type;
615 
616 		/* Header must be checked, and gso_segs computed. */
617 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
618 		skb_shinfo(skb)->gso_segs = 0;
619 	}
620 	return 0;
621 }
622 
623 static void macvtap_skb_to_vnet_hdr(struct macvtap_queue *q,
624 				    const struct sk_buff *skb,
625 				    struct virtio_net_hdr *vnet_hdr)
626 {
627 	memset(vnet_hdr, 0, sizeof(*vnet_hdr));
628 
629 	if (skb_is_gso(skb)) {
630 		struct skb_shared_info *sinfo = skb_shinfo(skb);
631 
632 		/* This is a hint as to how much should be linear. */
633 		vnet_hdr->hdr_len = cpu_to_macvtap16(q, skb_headlen(skb));
634 		vnet_hdr->gso_size = cpu_to_macvtap16(q, sinfo->gso_size);
635 		if (sinfo->gso_type & SKB_GSO_TCPV4)
636 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
637 		else if (sinfo->gso_type & SKB_GSO_TCPV6)
638 			vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
639 		else
640 			BUG();
641 		if (sinfo->gso_type & SKB_GSO_TCP_ECN)
642 			vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
643 	} else
644 		vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
645 
646 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
647 		vnet_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
648 		if (vlan_tx_tag_present(skb))
649 			vnet_hdr->csum_start = cpu_to_macvtap16(q,
650 				skb_checksum_start_offset(skb) + VLAN_HLEN);
651 		else
652 			vnet_hdr->csum_start = cpu_to_macvtap16(q,
653 				skb_checksum_start_offset(skb));
654 		vnet_hdr->csum_offset = cpu_to_macvtap16(q, skb->csum_offset);
655 	} else if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
656 		vnet_hdr->flags = VIRTIO_NET_HDR_F_DATA_VALID;
657 	} /* else everything is zero */
658 }
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(NET_IP_ALIGN);
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, NET_IP_ALIGN, 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 (vlan_tx_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(vlan_tx_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 
970 	/* tun/tap driver inverts the usage for TSO offloads, where
971 	 * setting the TSO bit means that the userspace wants to
972 	 * accept TSO frames and turning it off means that user space
973 	 * does not support TSO.
974 	 * For macvtap, we have to invert it to mean the same thing.
975 	 * When user space turns off TSO, we turn off GSO/LRO so that
976 	 * user-space will not receive TSO frames.
977 	 */
978 	if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6))
979 		features |= RX_OFFLOADS;
980 	else
981 		features &= ~RX_OFFLOADS;
982 
983 	/* tap_features are the same as features on tun/tap and
984 	 * reflect user expectations.
985 	 */
986 	vlan->tap_features = feature_mask;
987 	vlan->set_features = features;
988 	netdev_update_features(vlan->dev);
989 
990 	return 0;
991 }
992 
993 /*
994  * provide compatibility with generic tun/tap interface
995  */
996 static long macvtap_ioctl(struct file *file, unsigned int cmd,
997 			  unsigned long arg)
998 {
999 	struct macvtap_queue *q = file->private_data;
1000 	struct macvlan_dev *vlan;
1001 	void __user *argp = (void __user *)arg;
1002 	struct ifreq __user *ifr = argp;
1003 	unsigned int __user *up = argp;
1004 	unsigned short u;
1005 	int __user *sp = argp;
1006 	int s;
1007 	int ret;
1008 
1009 	switch (cmd) {
1010 	case TUNSETIFF:
1011 		/* ignore the name, just look at flags */
1012 		if (get_user(u, &ifr->ifr_flags))
1013 			return -EFAULT;
1014 
1015 		ret = 0;
1016 		if ((u & ~MACVTAP_FEATURES) != (IFF_NO_PI | IFF_TAP))
1017 			ret = -EINVAL;
1018 		else
1019 			q->flags = (q->flags & ~MACVTAP_FEATURES) | u;
1020 
1021 		return ret;
1022 
1023 	case TUNGETIFF:
1024 		rtnl_lock();
1025 		vlan = macvtap_get_vlan(q);
1026 		if (!vlan) {
1027 			rtnl_unlock();
1028 			return -ENOLINK;
1029 		}
1030 
1031 		ret = 0;
1032 		u = q->flags;
1033 		if (copy_to_user(&ifr->ifr_name, vlan->dev->name, IFNAMSIZ) ||
1034 		    put_user(u, &ifr->ifr_flags))
1035 			ret = -EFAULT;
1036 		macvtap_put_vlan(vlan);
1037 		rtnl_unlock();
1038 		return ret;
1039 
1040 	case TUNSETQUEUE:
1041 		if (get_user(u, &ifr->ifr_flags))
1042 			return -EFAULT;
1043 		rtnl_lock();
1044 		ret = macvtap_ioctl_set_queue(file, u);
1045 		rtnl_unlock();
1046 		return ret;
1047 
1048 	case TUNGETFEATURES:
1049 		if (put_user(IFF_TAP | IFF_NO_PI | MACVTAP_FEATURES, up))
1050 			return -EFAULT;
1051 		return 0;
1052 
1053 	case TUNSETSNDBUF:
1054 		if (get_user(u, up))
1055 			return -EFAULT;
1056 
1057 		q->sk.sk_sndbuf = u;
1058 		return 0;
1059 
1060 	case TUNGETVNETHDRSZ:
1061 		s = q->vnet_hdr_sz;
1062 		if (put_user(s, sp))
1063 			return -EFAULT;
1064 		return 0;
1065 
1066 	case TUNSETVNETHDRSZ:
1067 		if (get_user(s, sp))
1068 			return -EFAULT;
1069 		if (s < (int)sizeof(struct virtio_net_hdr))
1070 			return -EINVAL;
1071 
1072 		q->vnet_hdr_sz = s;
1073 		return 0;
1074 
1075 	case TUNGETVNETLE:
1076 		s = !!(q->flags & MACVTAP_VNET_LE);
1077 		if (put_user(s, sp))
1078 			return -EFAULT;
1079 		return 0;
1080 
1081 	case TUNSETVNETLE:
1082 		if (get_user(s, sp))
1083 			return -EFAULT;
1084 		if (s)
1085 			q->flags |= MACVTAP_VNET_LE;
1086 		else
1087 			q->flags &= ~MACVTAP_VNET_LE;
1088 		return 0;
1089 
1090 	case TUNSETOFFLOAD:
1091 		/* let the user check for future flags */
1092 		if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1093 			    TUN_F_TSO_ECN))
1094 			return -EINVAL;
1095 
1096 		rtnl_lock();
1097 		ret = set_offload(q, arg);
1098 		rtnl_unlock();
1099 		return ret;
1100 
1101 	default:
1102 		return -EINVAL;
1103 	}
1104 }
1105 
1106 #ifdef CONFIG_COMPAT
1107 static long macvtap_compat_ioctl(struct file *file, unsigned int cmd,
1108 				 unsigned long arg)
1109 {
1110 	return macvtap_ioctl(file, cmd, (unsigned long)compat_ptr(arg));
1111 }
1112 #endif
1113 
1114 static const struct file_operations macvtap_fops = {
1115 	.owner		= THIS_MODULE,
1116 	.open		= macvtap_open,
1117 	.release	= macvtap_release,
1118 	.read		= new_sync_read,
1119 	.write		= new_sync_write,
1120 	.read_iter	= macvtap_read_iter,
1121 	.write_iter	= macvtap_write_iter,
1122 	.poll		= macvtap_poll,
1123 	.llseek		= no_llseek,
1124 	.unlocked_ioctl	= macvtap_ioctl,
1125 #ifdef CONFIG_COMPAT
1126 	.compat_ioctl	= macvtap_compat_ioctl,
1127 #endif
1128 };
1129 
1130 static int macvtap_sendmsg(struct kiocb *iocb, struct socket *sock,
1131 			   struct msghdr *m, size_t total_len)
1132 {
1133 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1134 	return macvtap_get_user(q, m, &m->msg_iter, m->msg_flags & MSG_DONTWAIT);
1135 }
1136 
1137 static int macvtap_recvmsg(struct kiocb *iocb, struct socket *sock,
1138 			   struct msghdr *m, size_t total_len,
1139 			   int flags)
1140 {
1141 	struct macvtap_queue *q = container_of(sock, struct macvtap_queue, sock);
1142 	int ret;
1143 	if (flags & ~(MSG_DONTWAIT|MSG_TRUNC))
1144 		return -EINVAL;
1145 	ret = macvtap_do_read(q, &m->msg_iter, flags & MSG_DONTWAIT);
1146 	if (ret > total_len) {
1147 		m->msg_flags |= MSG_TRUNC;
1148 		ret = flags & MSG_TRUNC ? ret : total_len;
1149 	}
1150 	return ret;
1151 }
1152 
1153 /* Ops structure to mimic raw sockets with tun */
1154 static const struct proto_ops macvtap_socket_ops = {
1155 	.sendmsg = macvtap_sendmsg,
1156 	.recvmsg = macvtap_recvmsg,
1157 };
1158 
1159 /* Get an underlying socket object from tun file.  Returns error unless file is
1160  * attached to a device.  The returned object works like a packet socket, it
1161  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1162  * holding a reference to the file for as long as the socket is in use. */
1163 struct socket *macvtap_get_socket(struct file *file)
1164 {
1165 	struct macvtap_queue *q;
1166 	if (file->f_op != &macvtap_fops)
1167 		return ERR_PTR(-EINVAL);
1168 	q = file->private_data;
1169 	if (!q)
1170 		return ERR_PTR(-EBADFD);
1171 	return &q->sock;
1172 }
1173 EXPORT_SYMBOL_GPL(macvtap_get_socket);
1174 
1175 static int macvtap_device_event(struct notifier_block *unused,
1176 				unsigned long event, void *ptr)
1177 {
1178 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1179 	struct macvlan_dev *vlan;
1180 	struct device *classdev;
1181 	dev_t devt;
1182 	int err;
1183 
1184 	if (dev->rtnl_link_ops != &macvtap_link_ops)
1185 		return NOTIFY_DONE;
1186 
1187 	vlan = netdev_priv(dev);
1188 
1189 	switch (event) {
1190 	case NETDEV_REGISTER:
1191 		/* Create the device node here after the network device has
1192 		 * been registered but before register_netdevice has
1193 		 * finished running.
1194 		 */
1195 		err = macvtap_get_minor(vlan);
1196 		if (err)
1197 			return notifier_from_errno(err);
1198 
1199 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1200 		classdev = device_create(macvtap_class, &dev->dev, devt,
1201 					 dev, "tap%d", dev->ifindex);
1202 		if (IS_ERR(classdev)) {
1203 			macvtap_free_minor(vlan);
1204 			return notifier_from_errno(PTR_ERR(classdev));
1205 		}
1206 		break;
1207 	case NETDEV_UNREGISTER:
1208 		devt = MKDEV(MAJOR(macvtap_major), vlan->minor);
1209 		device_destroy(macvtap_class, devt);
1210 		macvtap_free_minor(vlan);
1211 		break;
1212 	}
1213 
1214 	return NOTIFY_DONE;
1215 }
1216 
1217 static struct notifier_block macvtap_notifier_block __read_mostly = {
1218 	.notifier_call	= macvtap_device_event,
1219 };
1220 
1221 static int macvtap_init(void)
1222 {
1223 	int err;
1224 
1225 	err = alloc_chrdev_region(&macvtap_major, 0,
1226 				MACVTAP_NUM_DEVS, "macvtap");
1227 	if (err)
1228 		goto out1;
1229 
1230 	cdev_init(&macvtap_cdev, &macvtap_fops);
1231 	err = cdev_add(&macvtap_cdev, macvtap_major, MACVTAP_NUM_DEVS);
1232 	if (err)
1233 		goto out2;
1234 
1235 	macvtap_class = class_create(THIS_MODULE, "macvtap");
1236 	if (IS_ERR(macvtap_class)) {
1237 		err = PTR_ERR(macvtap_class);
1238 		goto out3;
1239 	}
1240 
1241 	err = register_netdevice_notifier(&macvtap_notifier_block);
1242 	if (err)
1243 		goto out4;
1244 
1245 	err = macvlan_link_register(&macvtap_link_ops);
1246 	if (err)
1247 		goto out5;
1248 
1249 	return 0;
1250 
1251 out5:
1252 	unregister_netdevice_notifier(&macvtap_notifier_block);
1253 out4:
1254 	class_unregister(macvtap_class);
1255 out3:
1256 	cdev_del(&macvtap_cdev);
1257 out2:
1258 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1259 out1:
1260 	return err;
1261 }
1262 module_init(macvtap_init);
1263 
1264 static void macvtap_exit(void)
1265 {
1266 	rtnl_link_unregister(&macvtap_link_ops);
1267 	unregister_netdevice_notifier(&macvtap_notifier_block);
1268 	class_unregister(macvtap_class);
1269 	cdev_del(&macvtap_cdev);
1270 	unregister_chrdev_region(macvtap_major, MACVTAP_NUM_DEVS);
1271 }
1272 module_exit(macvtap_exit);
1273 
1274 MODULE_ALIAS_RTNL_LINK("macvtap");
1275 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1276 MODULE_LICENSE("GPL");
1277