xref: /linux/drivers/net/tap.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/etherdevice.h>
3 #include <linux/if_tap.h>
4 #include <linux/if_vlan.h>
5 #include <linux/interrupt.h>
6 #include <linux/nsproxy.h>
7 #include <linux/compat.h>
8 #include <linux/if_tun.h>
9 #include <linux/module.h>
10 #include <linux/skbuff.h>
11 #include <linux/cache.h>
12 #include <linux/sched/signal.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/wait.h>
16 #include <linux/cdev.h>
17 #include <linux/idr.h>
18 #include <linux/fs.h>
19 #include <linux/uio.h>
20 
21 #include <net/net_namespace.h>
22 #include <net/rtnetlink.h>
23 #include <net/sock.h>
24 #include <linux/virtio_net.h>
25 #include <linux/skb_array.h>
26 
27 #define TAP_IFFEATURES (IFF_VNET_HDR | IFF_MULTI_QUEUE)
28 
29 #define TAP_VNET_LE 0x80000000
30 #define TAP_VNET_BE 0x40000000
31 
32 #ifdef CONFIG_TUN_VNET_CROSS_LE
33 static inline bool tap_legacy_is_little_endian(struct tap_queue *q)
34 {
35 	return q->flags & TAP_VNET_BE ? false :
36 		virtio_legacy_is_little_endian();
37 }
38 
39 static long tap_get_vnet_be(struct tap_queue *q, int __user *sp)
40 {
41 	int s = !!(q->flags & TAP_VNET_BE);
42 
43 	if (put_user(s, sp))
44 		return -EFAULT;
45 
46 	return 0;
47 }
48 
49 static long tap_set_vnet_be(struct tap_queue *q, int __user *sp)
50 {
51 	int s;
52 
53 	if (get_user(s, sp))
54 		return -EFAULT;
55 
56 	if (s)
57 		q->flags |= TAP_VNET_BE;
58 	else
59 		q->flags &= ~TAP_VNET_BE;
60 
61 	return 0;
62 }
63 #else
64 static inline bool tap_legacy_is_little_endian(struct tap_queue *q)
65 {
66 	return virtio_legacy_is_little_endian();
67 }
68 
69 static long tap_get_vnet_be(struct tap_queue *q, int __user *argp)
70 {
71 	return -EINVAL;
72 }
73 
74 static long tap_set_vnet_be(struct tap_queue *q, int __user *argp)
75 {
76 	return -EINVAL;
77 }
78 #endif /* CONFIG_TUN_VNET_CROSS_LE */
79 
80 static inline bool tap_is_little_endian(struct tap_queue *q)
81 {
82 	return q->flags & TAP_VNET_LE ||
83 		tap_legacy_is_little_endian(q);
84 }
85 
86 static inline u16 tap16_to_cpu(struct tap_queue *q, __virtio16 val)
87 {
88 	return __virtio16_to_cpu(tap_is_little_endian(q), val);
89 }
90 
91 static inline __virtio16 cpu_to_tap16(struct tap_queue *q, u16 val)
92 {
93 	return __cpu_to_virtio16(tap_is_little_endian(q), val);
94 }
95 
96 static struct proto tap_proto = {
97 	.name = "tap",
98 	.owner = THIS_MODULE,
99 	.obj_size = sizeof(struct tap_queue),
100 };
101 
102 #define TAP_NUM_DEVS (1U << MINORBITS)
103 
104 static LIST_HEAD(major_list);
105 
106 struct major_info {
107 	struct rcu_head rcu;
108 	dev_t major;
109 	struct idr minor_idr;
110 	spinlock_t minor_lock;
111 	const char *device_name;
112 	struct list_head next;
113 };
114 
115 #define GOODCOPY_LEN 128
116 
117 static const struct proto_ops tap_socket_ops;
118 
119 #define RX_OFFLOADS (NETIF_F_GRO | NETIF_F_LRO)
120 #define TAP_FEATURES (NETIF_F_GSO | NETIF_F_SG | NETIF_F_FRAGLIST)
121 
122 static struct tap_dev *tap_dev_get_rcu(const struct net_device *dev)
123 {
124 	return rcu_dereference(dev->rx_handler_data);
125 }
126 
127 /*
128  * RCU usage:
129  * The tap_queue and the macvlan_dev are loosely coupled, the
130  * pointers from one to the other can only be read while rcu_read_lock
131  * or rtnl is held.
132  *
133  * Both the file and the macvlan_dev hold a reference on the tap_queue
134  * through sock_hold(&q->sk). When the macvlan_dev goes away first,
135  * q->vlan becomes inaccessible. When the files gets closed,
136  * tap_get_queue() fails.
137  *
138  * There may still be references to the struct sock inside of the
139  * queue from outbound SKBs, but these never reference back to the
140  * file or the dev. The data structure is freed through __sk_free
141  * when both our references and any pending SKBs are gone.
142  */
143 
144 static int tap_enable_queue(struct tap_dev *tap, struct file *file,
145 			    struct tap_queue *q)
146 {
147 	int err = -EINVAL;
148 
149 	ASSERT_RTNL();
150 
151 	if (q->enabled)
152 		goto out;
153 
154 	err = 0;
155 	rcu_assign_pointer(tap->taps[tap->numvtaps], q);
156 	q->queue_index = tap->numvtaps;
157 	q->enabled = true;
158 
159 	tap->numvtaps++;
160 out:
161 	return err;
162 }
163 
164 /* Requires RTNL */
165 static int tap_set_queue(struct tap_dev *tap, struct file *file,
166 			 struct tap_queue *q)
167 {
168 	if (tap->numqueues == MAX_TAP_QUEUES)
169 		return -EBUSY;
170 
171 	rcu_assign_pointer(q->tap, tap);
172 	rcu_assign_pointer(tap->taps[tap->numvtaps], q);
173 	sock_hold(&q->sk);
174 
175 	q->file = file;
176 	q->queue_index = tap->numvtaps;
177 	q->enabled = true;
178 	file->private_data = q;
179 	list_add_tail(&q->next, &tap->queue_list);
180 
181 	tap->numvtaps++;
182 	tap->numqueues++;
183 
184 	return 0;
185 }
186 
187 static int tap_disable_queue(struct tap_queue *q)
188 {
189 	struct tap_dev *tap;
190 	struct tap_queue *nq;
191 
192 	ASSERT_RTNL();
193 	if (!q->enabled)
194 		return -EINVAL;
195 
196 	tap = rtnl_dereference(q->tap);
197 
198 	if (tap) {
199 		int index = q->queue_index;
200 		BUG_ON(index >= tap->numvtaps);
201 		nq = rtnl_dereference(tap->taps[tap->numvtaps - 1]);
202 		nq->queue_index = index;
203 
204 		rcu_assign_pointer(tap->taps[index], nq);
205 		RCU_INIT_POINTER(tap->taps[tap->numvtaps - 1], NULL);
206 		q->enabled = false;
207 
208 		tap->numvtaps--;
209 	}
210 
211 	return 0;
212 }
213 
214 /*
215  * The file owning the queue got closed, give up both
216  * the reference that the files holds as well as the
217  * one from the macvlan_dev if that still exists.
218  *
219  * Using the spinlock makes sure that we don't get
220  * to the queue again after destroying it.
221  */
222 static void tap_put_queue(struct tap_queue *q)
223 {
224 	struct tap_dev *tap;
225 
226 	rtnl_lock();
227 	tap = rtnl_dereference(q->tap);
228 
229 	if (tap) {
230 		if (q->enabled)
231 			BUG_ON(tap_disable_queue(q));
232 
233 		tap->numqueues--;
234 		RCU_INIT_POINTER(q->tap, NULL);
235 		sock_put(&q->sk);
236 		list_del_init(&q->next);
237 	}
238 
239 	rtnl_unlock();
240 
241 	synchronize_rcu();
242 	sock_put(&q->sk);
243 }
244 
245 /*
246  * Select a queue based on the rxq of the device on which this packet
247  * arrived. If the incoming device is not mq, calculate a flow hash
248  * to select a queue. If all fails, find the first available queue.
249  * Cache vlan->numvtaps since it can become zero during the execution
250  * of this function.
251  */
252 static struct tap_queue *tap_get_queue(struct tap_dev *tap,
253 				       struct sk_buff *skb)
254 {
255 	struct tap_queue *queue = NULL;
256 	/* Access to taps array is protected by rcu, but access to numvtaps
257 	 * isn't. Below we use it to lookup a queue, but treat it as a hint
258 	 * and validate that the result isn't NULL - in case we are
259 	 * racing against queue removal.
260 	 */
261 	int numvtaps = READ_ONCE(tap->numvtaps);
262 	__u32 rxq;
263 
264 	if (!numvtaps)
265 		goto out;
266 
267 	if (numvtaps == 1)
268 		goto single;
269 
270 	/* Check if we can use flow to select a queue */
271 	rxq = skb_get_hash(skb);
272 	if (rxq) {
273 		queue = rcu_dereference(tap->taps[rxq % numvtaps]);
274 		goto out;
275 	}
276 
277 	if (likely(skb_rx_queue_recorded(skb))) {
278 		rxq = skb_get_rx_queue(skb);
279 
280 		while (unlikely(rxq >= numvtaps))
281 			rxq -= numvtaps;
282 
283 		queue = rcu_dereference(tap->taps[rxq]);
284 		goto out;
285 	}
286 
287 single:
288 	queue = rcu_dereference(tap->taps[0]);
289 out:
290 	return queue;
291 }
292 
293 /*
294  * The net_device is going away, give up the reference
295  * that it holds on all queues and safely set the pointer
296  * from the queues to NULL.
297  */
298 void tap_del_queues(struct tap_dev *tap)
299 {
300 	struct tap_queue *q, *tmp;
301 
302 	ASSERT_RTNL();
303 	list_for_each_entry_safe(q, tmp, &tap->queue_list, next) {
304 		list_del_init(&q->next);
305 		RCU_INIT_POINTER(q->tap, NULL);
306 		if (q->enabled)
307 			tap->numvtaps--;
308 		tap->numqueues--;
309 		sock_put(&q->sk);
310 	}
311 	BUG_ON(tap->numvtaps);
312 	BUG_ON(tap->numqueues);
313 	/* guarantee that any future tap_set_queue will fail */
314 	tap->numvtaps = MAX_TAP_QUEUES;
315 }
316 EXPORT_SYMBOL_GPL(tap_del_queues);
317 
318 rx_handler_result_t tap_handle_frame(struct sk_buff **pskb)
319 {
320 	struct sk_buff *skb = *pskb;
321 	struct net_device *dev = skb->dev;
322 	struct tap_dev *tap;
323 	struct tap_queue *q;
324 	netdev_features_t features = TAP_FEATURES;
325 
326 	tap = tap_dev_get_rcu(dev);
327 	if (!tap)
328 		return RX_HANDLER_PASS;
329 
330 	q = tap_get_queue(tap, skb);
331 	if (!q)
332 		return RX_HANDLER_PASS;
333 
334 	skb_push(skb, ETH_HLEN);
335 
336 	/* Apply the forward feature mask so that we perform segmentation
337 	 * according to users wishes.  This only works if VNET_HDR is
338 	 * enabled.
339 	 */
340 	if (q->flags & IFF_VNET_HDR)
341 		features |= tap->tap_features;
342 	if (netif_needs_gso(skb, features)) {
343 		struct sk_buff *segs = __skb_gso_segment(skb, features, false);
344 
345 		if (IS_ERR(segs))
346 			goto drop;
347 
348 		if (!segs) {
349 			if (ptr_ring_produce(&q->ring, skb))
350 				goto drop;
351 			goto wake_up;
352 		}
353 
354 		consume_skb(skb);
355 		while (segs) {
356 			struct sk_buff *nskb = segs->next;
357 
358 			segs->next = NULL;
359 			if (ptr_ring_produce(&q->ring, segs)) {
360 				kfree_skb(segs);
361 				kfree_skb_list(nskb);
362 				break;
363 			}
364 			segs = nskb;
365 		}
366 	} else {
367 		/* If we receive a partial checksum and the tap side
368 		 * doesn't support checksum offload, compute the checksum.
369 		 * Note: it doesn't matter which checksum feature to
370 		 *	  check, we either support them all or none.
371 		 */
372 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
373 		    !(features & NETIF_F_CSUM_MASK) &&
374 		    skb_checksum_help(skb))
375 			goto drop;
376 		if (ptr_ring_produce(&q->ring, skb))
377 			goto drop;
378 	}
379 
380 wake_up:
381 	wake_up_interruptible_poll(sk_sleep(&q->sk), EPOLLIN | EPOLLRDNORM | EPOLLRDBAND);
382 	return RX_HANDLER_CONSUMED;
383 
384 drop:
385 	/* Count errors/drops only here, thus don't care about args. */
386 	if (tap->count_rx_dropped)
387 		tap->count_rx_dropped(tap);
388 	kfree_skb(skb);
389 	return RX_HANDLER_CONSUMED;
390 }
391 EXPORT_SYMBOL_GPL(tap_handle_frame);
392 
393 static struct major_info *tap_get_major(int major)
394 {
395 	struct major_info *tap_major;
396 
397 	list_for_each_entry_rcu(tap_major, &major_list, next) {
398 		if (tap_major->major == major)
399 			return tap_major;
400 	}
401 
402 	return NULL;
403 }
404 
405 int tap_get_minor(dev_t major, struct tap_dev *tap)
406 {
407 	int retval = -ENOMEM;
408 	struct major_info *tap_major;
409 
410 	rcu_read_lock();
411 	tap_major = tap_get_major(MAJOR(major));
412 	if (!tap_major) {
413 		retval = -EINVAL;
414 		goto unlock;
415 	}
416 
417 	spin_lock(&tap_major->minor_lock);
418 	retval = idr_alloc(&tap_major->minor_idr, tap, 1, TAP_NUM_DEVS, GFP_ATOMIC);
419 	if (retval >= 0) {
420 		tap->minor = retval;
421 	} else if (retval == -ENOSPC) {
422 		netdev_err(tap->dev, "Too many tap devices\n");
423 		retval = -EINVAL;
424 	}
425 	spin_unlock(&tap_major->minor_lock);
426 
427 unlock:
428 	rcu_read_unlock();
429 	return retval < 0 ? retval : 0;
430 }
431 EXPORT_SYMBOL_GPL(tap_get_minor);
432 
433 void tap_free_minor(dev_t major, struct tap_dev *tap)
434 {
435 	struct major_info *tap_major;
436 
437 	rcu_read_lock();
438 	tap_major = tap_get_major(MAJOR(major));
439 	if (!tap_major) {
440 		goto unlock;
441 	}
442 
443 	spin_lock(&tap_major->minor_lock);
444 	if (tap->minor) {
445 		idr_remove(&tap_major->minor_idr, tap->minor);
446 		tap->minor = 0;
447 	}
448 	spin_unlock(&tap_major->minor_lock);
449 
450 unlock:
451 	rcu_read_unlock();
452 }
453 EXPORT_SYMBOL_GPL(tap_free_minor);
454 
455 static struct tap_dev *dev_get_by_tap_file(int major, int minor)
456 {
457 	struct net_device *dev = NULL;
458 	struct tap_dev *tap;
459 	struct major_info *tap_major;
460 
461 	rcu_read_lock();
462 	tap_major = tap_get_major(major);
463 	if (!tap_major) {
464 		tap = NULL;
465 		goto unlock;
466 	}
467 
468 	spin_lock(&tap_major->minor_lock);
469 	tap = idr_find(&tap_major->minor_idr, minor);
470 	if (tap) {
471 		dev = tap->dev;
472 		dev_hold(dev);
473 	}
474 	spin_unlock(&tap_major->minor_lock);
475 
476 unlock:
477 	rcu_read_unlock();
478 	return tap;
479 }
480 
481 static void tap_sock_write_space(struct sock *sk)
482 {
483 	wait_queue_head_t *wqueue;
484 
485 	if (!sock_writeable(sk) ||
486 	    !test_and_clear_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags))
487 		return;
488 
489 	wqueue = sk_sleep(sk);
490 	if (wqueue && waitqueue_active(wqueue))
491 		wake_up_interruptible_poll(wqueue, EPOLLOUT | EPOLLWRNORM | EPOLLWRBAND);
492 }
493 
494 static void tap_sock_destruct(struct sock *sk)
495 {
496 	struct tap_queue *q = container_of(sk, struct tap_queue, sk);
497 
498 	ptr_ring_cleanup(&q->ring, __skb_array_destroy_skb);
499 }
500 
501 static int tap_open(struct inode *inode, struct file *file)
502 {
503 	struct net *net = current->nsproxy->net_ns;
504 	struct tap_dev *tap;
505 	struct tap_queue *q;
506 	int err = -ENODEV;
507 
508 	rtnl_lock();
509 	tap = dev_get_by_tap_file(imajor(inode), iminor(inode));
510 	if (!tap)
511 		goto err;
512 
513 	err = -ENOMEM;
514 	q = (struct tap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
515 					     &tap_proto, 0);
516 	if (!q)
517 		goto err;
518 	if (ptr_ring_init(&q->ring, tap->dev->tx_queue_len, GFP_KERNEL)) {
519 		sk_free(&q->sk);
520 		goto err;
521 	}
522 
523 	RCU_INIT_POINTER(q->sock.wq, &q->wq);
524 	init_waitqueue_head(&q->wq.wait);
525 	q->sock.type = SOCK_RAW;
526 	q->sock.state = SS_CONNECTED;
527 	q->sock.file = file;
528 	q->sock.ops = &tap_socket_ops;
529 	sock_init_data(&q->sock, &q->sk);
530 	q->sk.sk_write_space = tap_sock_write_space;
531 	q->sk.sk_destruct = tap_sock_destruct;
532 	q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
533 	q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
534 
535 	/*
536 	 * so far only KVM virtio_net uses tap, enable zero copy between
537 	 * guest kernel and host kernel when lower device supports zerocopy
538 	 *
539 	 * The macvlan supports zerocopy iff the lower device supports zero
540 	 * copy so we don't have to look at the lower device directly.
541 	 */
542 	if ((tap->dev->features & NETIF_F_HIGHDMA) && (tap->dev->features & NETIF_F_SG))
543 		sock_set_flag(&q->sk, SOCK_ZEROCOPY);
544 
545 	err = tap_set_queue(tap, file, q);
546 	if (err) {
547 		/* tap_sock_destruct() will take care of freeing ptr_ring */
548 		goto err_put;
549 	}
550 
551 	dev_put(tap->dev);
552 
553 	rtnl_unlock();
554 	return err;
555 
556 err_put:
557 	sock_put(&q->sk);
558 err:
559 	if (tap)
560 		dev_put(tap->dev);
561 
562 	rtnl_unlock();
563 	return err;
564 }
565 
566 static int tap_release(struct inode *inode, struct file *file)
567 {
568 	struct tap_queue *q = file->private_data;
569 	tap_put_queue(q);
570 	return 0;
571 }
572 
573 static __poll_t tap_poll(struct file *file, poll_table *wait)
574 {
575 	struct tap_queue *q = file->private_data;
576 	__poll_t mask = EPOLLERR;
577 
578 	if (!q)
579 		goto out;
580 
581 	mask = 0;
582 	poll_wait(file, &q->wq.wait, wait);
583 
584 	if (!ptr_ring_empty(&q->ring))
585 		mask |= EPOLLIN | EPOLLRDNORM;
586 
587 	if (sock_writeable(&q->sk) ||
588 	    (!test_and_set_bit(SOCKWQ_ASYNC_NOSPACE, &q->sock.flags) &&
589 	     sock_writeable(&q->sk)))
590 		mask |= EPOLLOUT | EPOLLWRNORM;
591 
592 out:
593 	return mask;
594 }
595 
596 static inline struct sk_buff *tap_alloc_skb(struct sock *sk, size_t prepad,
597 					    size_t len, size_t linear,
598 						int noblock, int *err)
599 {
600 	struct sk_buff *skb;
601 
602 	/* Under a page?  Don't bother with paged skb. */
603 	if (prepad + len < PAGE_SIZE || !linear)
604 		linear = len;
605 
606 	skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
607 				   err, 0);
608 	if (!skb)
609 		return NULL;
610 
611 	skb_reserve(skb, prepad);
612 	skb_put(skb, linear);
613 	skb->data_len = len - linear;
614 	skb->len += len - linear;
615 
616 	return skb;
617 }
618 
619 /* Neighbour code has some assumptions on HH_DATA_MOD alignment */
620 #define TAP_RESERVE HH_DATA_OFF(ETH_HLEN)
621 
622 /* Get packet from user space buffer */
623 static ssize_t tap_get_user(struct tap_queue *q, void *msg_control,
624 			    struct iov_iter *from, int noblock)
625 {
626 	int good_linear = SKB_MAX_HEAD(TAP_RESERVE);
627 	struct sk_buff *skb;
628 	struct tap_dev *tap;
629 	unsigned long total_len = iov_iter_count(from);
630 	unsigned long len = total_len;
631 	int err;
632 	struct virtio_net_hdr vnet_hdr = { 0 };
633 	int vnet_hdr_len = 0;
634 	int copylen = 0;
635 	int depth;
636 	bool zerocopy = false;
637 	size_t linear;
638 
639 	if (q->flags & IFF_VNET_HDR) {
640 		vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
641 
642 		err = -EINVAL;
643 		if (len < vnet_hdr_len)
644 			goto err;
645 		len -= vnet_hdr_len;
646 
647 		err = -EFAULT;
648 		if (!copy_from_iter_full(&vnet_hdr, sizeof(vnet_hdr), from))
649 			goto err;
650 		iov_iter_advance(from, vnet_hdr_len - sizeof(vnet_hdr));
651 		if ((vnet_hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
652 		     tap16_to_cpu(q, vnet_hdr.csum_start) +
653 		     tap16_to_cpu(q, vnet_hdr.csum_offset) + 2 >
654 			     tap16_to_cpu(q, vnet_hdr.hdr_len))
655 			vnet_hdr.hdr_len = cpu_to_tap16(q,
656 				 tap16_to_cpu(q, vnet_hdr.csum_start) +
657 				 tap16_to_cpu(q, vnet_hdr.csum_offset) + 2);
658 		err = -EINVAL;
659 		if (tap16_to_cpu(q, vnet_hdr.hdr_len) > len)
660 			goto err;
661 	}
662 
663 	err = -EINVAL;
664 	if (unlikely(len < ETH_HLEN))
665 		goto err;
666 
667 	if (msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
668 		struct iov_iter i;
669 
670 		copylen = vnet_hdr.hdr_len ?
671 			tap16_to_cpu(q, vnet_hdr.hdr_len) : GOODCOPY_LEN;
672 		if (copylen > good_linear)
673 			copylen = good_linear;
674 		else if (copylen < ETH_HLEN)
675 			copylen = ETH_HLEN;
676 		linear = copylen;
677 		i = *from;
678 		iov_iter_advance(&i, copylen);
679 		if (iov_iter_npages(&i, INT_MAX) <= MAX_SKB_FRAGS)
680 			zerocopy = true;
681 	}
682 
683 	if (!zerocopy) {
684 		copylen = len;
685 		linear = tap16_to_cpu(q, vnet_hdr.hdr_len);
686 		if (linear > good_linear)
687 			linear = good_linear;
688 		else if (linear < ETH_HLEN)
689 			linear = ETH_HLEN;
690 	}
691 
692 	skb = tap_alloc_skb(&q->sk, TAP_RESERVE, copylen,
693 			    linear, noblock, &err);
694 	if (!skb)
695 		goto err;
696 
697 	if (zerocopy)
698 		err = zerocopy_sg_from_iter(skb, from);
699 	else
700 		err = skb_copy_datagram_from_iter(skb, 0, from, len);
701 
702 	if (err)
703 		goto err_kfree;
704 
705 	skb_set_network_header(skb, ETH_HLEN);
706 	skb_reset_mac_header(skb);
707 	skb->protocol = eth_hdr(skb)->h_proto;
708 
709 	if (vnet_hdr_len) {
710 		err = virtio_net_hdr_to_skb(skb, &vnet_hdr,
711 					    tap_is_little_endian(q));
712 		if (err)
713 			goto err_kfree;
714 	}
715 
716 	skb_probe_transport_header(skb);
717 
718 	/* Move network header to the right position for VLAN tagged packets */
719 	if ((skb->protocol == htons(ETH_P_8021Q) ||
720 	     skb->protocol == htons(ETH_P_8021AD)) &&
721 	    __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
722 		skb_set_network_header(skb, depth);
723 
724 	rcu_read_lock();
725 	tap = rcu_dereference(q->tap);
726 	/* copy skb_ubuf_info for callback when skb has no error */
727 	if (zerocopy) {
728 		skb_shinfo(skb)->destructor_arg = msg_control;
729 		skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
730 		skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
731 	} else if (msg_control) {
732 		struct ubuf_info *uarg = msg_control;
733 		uarg->callback(uarg, false);
734 	}
735 
736 	if (tap) {
737 		skb->dev = tap->dev;
738 		dev_queue_xmit(skb);
739 	} else {
740 		kfree_skb(skb);
741 	}
742 	rcu_read_unlock();
743 
744 	return total_len;
745 
746 err_kfree:
747 	kfree_skb(skb);
748 
749 err:
750 	rcu_read_lock();
751 	tap = rcu_dereference(q->tap);
752 	if (tap && tap->count_tx_dropped)
753 		tap->count_tx_dropped(tap);
754 	rcu_read_unlock();
755 
756 	return err;
757 }
758 
759 static ssize_t tap_write_iter(struct kiocb *iocb, struct iov_iter *from)
760 {
761 	struct file *file = iocb->ki_filp;
762 	struct tap_queue *q = file->private_data;
763 
764 	return tap_get_user(q, NULL, from, file->f_flags & O_NONBLOCK);
765 }
766 
767 /* Put packet to the user space buffer */
768 static ssize_t tap_put_user(struct tap_queue *q,
769 			    const struct sk_buff *skb,
770 			    struct iov_iter *iter)
771 {
772 	int ret;
773 	int vnet_hdr_len = 0;
774 	int vlan_offset = 0;
775 	int total;
776 
777 	if (q->flags & IFF_VNET_HDR) {
778 		int vlan_hlen = skb_vlan_tag_present(skb) ? VLAN_HLEN : 0;
779 		struct virtio_net_hdr vnet_hdr;
780 
781 		vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
782 		if (iov_iter_count(iter) < vnet_hdr_len)
783 			return -EINVAL;
784 
785 		if (virtio_net_hdr_from_skb(skb, &vnet_hdr,
786 					    tap_is_little_endian(q), true,
787 					    vlan_hlen))
788 			BUG();
789 
790 		if (copy_to_iter(&vnet_hdr, sizeof(vnet_hdr), iter) !=
791 		    sizeof(vnet_hdr))
792 			return -EFAULT;
793 
794 		iov_iter_advance(iter, vnet_hdr_len - sizeof(vnet_hdr));
795 	}
796 	total = vnet_hdr_len;
797 	total += skb->len;
798 
799 	if (skb_vlan_tag_present(skb)) {
800 		struct {
801 			__be16 h_vlan_proto;
802 			__be16 h_vlan_TCI;
803 		} veth;
804 		veth.h_vlan_proto = skb->vlan_proto;
805 		veth.h_vlan_TCI = htons(skb_vlan_tag_get(skb));
806 
807 		vlan_offset = offsetof(struct vlan_ethhdr, h_vlan_proto);
808 		total += VLAN_HLEN;
809 
810 		ret = skb_copy_datagram_iter(skb, 0, iter, vlan_offset);
811 		if (ret || !iov_iter_count(iter))
812 			goto done;
813 
814 		ret = copy_to_iter(&veth, sizeof(veth), iter);
815 		if (ret != sizeof(veth) || !iov_iter_count(iter))
816 			goto done;
817 	}
818 
819 	ret = skb_copy_datagram_iter(skb, vlan_offset, iter,
820 				     skb->len - vlan_offset);
821 
822 done:
823 	return ret ? ret : total;
824 }
825 
826 static ssize_t tap_do_read(struct tap_queue *q,
827 			   struct iov_iter *to,
828 			   int noblock, struct sk_buff *skb)
829 {
830 	DEFINE_WAIT(wait);
831 	ssize_t ret = 0;
832 
833 	if (!iov_iter_count(to)) {
834 		kfree_skb(skb);
835 		return 0;
836 	}
837 
838 	if (skb)
839 		goto put;
840 
841 	while (1) {
842 		if (!noblock)
843 			prepare_to_wait(sk_sleep(&q->sk), &wait,
844 					TASK_INTERRUPTIBLE);
845 
846 		/* Read frames from the queue */
847 		skb = ptr_ring_consume(&q->ring);
848 		if (skb)
849 			break;
850 		if (noblock) {
851 			ret = -EAGAIN;
852 			break;
853 		}
854 		if (signal_pending(current)) {
855 			ret = -ERESTARTSYS;
856 			break;
857 		}
858 		/* Nothing to read, let's sleep */
859 		schedule();
860 	}
861 	if (!noblock)
862 		finish_wait(sk_sleep(&q->sk), &wait);
863 
864 put:
865 	if (skb) {
866 		ret = tap_put_user(q, skb, to);
867 		if (unlikely(ret < 0))
868 			kfree_skb(skb);
869 		else
870 			consume_skb(skb);
871 	}
872 	return ret;
873 }
874 
875 static ssize_t tap_read_iter(struct kiocb *iocb, struct iov_iter *to)
876 {
877 	struct file *file = iocb->ki_filp;
878 	struct tap_queue *q = file->private_data;
879 	ssize_t len = iov_iter_count(to), ret;
880 
881 	ret = tap_do_read(q, to, file->f_flags & O_NONBLOCK, NULL);
882 	ret = min_t(ssize_t, ret, len);
883 	if (ret > 0)
884 		iocb->ki_pos = ret;
885 	return ret;
886 }
887 
888 static struct tap_dev *tap_get_tap_dev(struct tap_queue *q)
889 {
890 	struct tap_dev *tap;
891 
892 	ASSERT_RTNL();
893 	tap = rtnl_dereference(q->tap);
894 	if (tap)
895 		dev_hold(tap->dev);
896 
897 	return tap;
898 }
899 
900 static void tap_put_tap_dev(struct tap_dev *tap)
901 {
902 	dev_put(tap->dev);
903 }
904 
905 static int tap_ioctl_set_queue(struct file *file, unsigned int flags)
906 {
907 	struct tap_queue *q = file->private_data;
908 	struct tap_dev *tap;
909 	int ret;
910 
911 	tap = tap_get_tap_dev(q);
912 	if (!tap)
913 		return -EINVAL;
914 
915 	if (flags & IFF_ATTACH_QUEUE)
916 		ret = tap_enable_queue(tap, file, q);
917 	else if (flags & IFF_DETACH_QUEUE)
918 		ret = tap_disable_queue(q);
919 	else
920 		ret = -EINVAL;
921 
922 	tap_put_tap_dev(tap);
923 	return ret;
924 }
925 
926 static int set_offload(struct tap_queue *q, unsigned long arg)
927 {
928 	struct tap_dev *tap;
929 	netdev_features_t features;
930 	netdev_features_t feature_mask = 0;
931 
932 	tap = rtnl_dereference(q->tap);
933 	if (!tap)
934 		return -ENOLINK;
935 
936 	features = tap->dev->features;
937 
938 	if (arg & TUN_F_CSUM) {
939 		feature_mask = NETIF_F_HW_CSUM;
940 
941 		if (arg & (TUN_F_TSO4 | TUN_F_TSO6)) {
942 			if (arg & TUN_F_TSO_ECN)
943 				feature_mask |= NETIF_F_TSO_ECN;
944 			if (arg & TUN_F_TSO4)
945 				feature_mask |= NETIF_F_TSO;
946 			if (arg & TUN_F_TSO6)
947 				feature_mask |= NETIF_F_TSO6;
948 		}
949 	}
950 
951 	/* tun/tap driver inverts the usage for TSO offloads, where
952 	 * setting the TSO bit means that the userspace wants to
953 	 * accept TSO frames and turning it off means that user space
954 	 * does not support TSO.
955 	 * For tap, we have to invert it to mean the same thing.
956 	 * When user space turns off TSO, we turn off GSO/LRO so that
957 	 * user-space will not receive TSO frames.
958 	 */
959 	if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6))
960 		features |= RX_OFFLOADS;
961 	else
962 		features &= ~RX_OFFLOADS;
963 
964 	/* tap_features are the same as features on tun/tap and
965 	 * reflect user expectations.
966 	 */
967 	tap->tap_features = feature_mask;
968 	if (tap->update_features)
969 		tap->update_features(tap, features);
970 
971 	return 0;
972 }
973 
974 /*
975  * provide compatibility with generic tun/tap interface
976  */
977 static long tap_ioctl(struct file *file, unsigned int cmd,
978 		      unsigned long arg)
979 {
980 	struct tap_queue *q = file->private_data;
981 	struct tap_dev *tap;
982 	void __user *argp = (void __user *)arg;
983 	struct ifreq __user *ifr = argp;
984 	unsigned int __user *up = argp;
985 	unsigned short u;
986 	int __user *sp = argp;
987 	struct sockaddr sa;
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 & ~TAP_IFFEATURES) != (IFF_NO_PI | IFF_TAP))
999 			ret = -EINVAL;
1000 		else
1001 			q->flags = (q->flags & ~TAP_IFFEATURES) | u;
1002 
1003 		return ret;
1004 
1005 	case TUNGETIFF:
1006 		rtnl_lock();
1007 		tap = tap_get_tap_dev(q);
1008 		if (!tap) {
1009 			rtnl_unlock();
1010 			return -ENOLINK;
1011 		}
1012 
1013 		ret = 0;
1014 		u = q->flags;
1015 		if (copy_to_user(&ifr->ifr_name, tap->dev->name, IFNAMSIZ) ||
1016 		    put_user(u, &ifr->ifr_flags))
1017 			ret = -EFAULT;
1018 		tap_put_tap_dev(tap);
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 = tap_ioctl_set_queue(file, u);
1027 		rtnl_unlock();
1028 		return ret;
1029 
1030 	case TUNGETFEATURES:
1031 		if (put_user(IFF_TAP | IFF_NO_PI | TAP_IFFEATURES, up))
1032 			return -EFAULT;
1033 		return 0;
1034 
1035 	case TUNSETSNDBUF:
1036 		if (get_user(s, sp))
1037 			return -EFAULT;
1038 		if (s <= 0)
1039 			return -EINVAL;
1040 
1041 		q->sk.sk_sndbuf = s;
1042 		return 0;
1043 
1044 	case TUNGETVNETHDRSZ:
1045 		s = q->vnet_hdr_sz;
1046 		if (put_user(s, sp))
1047 			return -EFAULT;
1048 		return 0;
1049 
1050 	case TUNSETVNETHDRSZ:
1051 		if (get_user(s, sp))
1052 			return -EFAULT;
1053 		if (s < (int)sizeof(struct virtio_net_hdr))
1054 			return -EINVAL;
1055 
1056 		q->vnet_hdr_sz = s;
1057 		return 0;
1058 
1059 	case TUNGETVNETLE:
1060 		s = !!(q->flags & TAP_VNET_LE);
1061 		if (put_user(s, sp))
1062 			return -EFAULT;
1063 		return 0;
1064 
1065 	case TUNSETVNETLE:
1066 		if (get_user(s, sp))
1067 			return -EFAULT;
1068 		if (s)
1069 			q->flags |= TAP_VNET_LE;
1070 		else
1071 			q->flags &= ~TAP_VNET_LE;
1072 		return 0;
1073 
1074 	case TUNGETVNETBE:
1075 		return tap_get_vnet_be(q, sp);
1076 
1077 	case TUNSETVNETBE:
1078 		return tap_set_vnet_be(q, sp);
1079 
1080 	case TUNSETOFFLOAD:
1081 		/* let the user check for future flags */
1082 		if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1083 			    TUN_F_TSO_ECN | TUN_F_UFO))
1084 			return -EINVAL;
1085 
1086 		rtnl_lock();
1087 		ret = set_offload(q, arg);
1088 		rtnl_unlock();
1089 		return ret;
1090 
1091 	case SIOCGIFHWADDR:
1092 		rtnl_lock();
1093 		tap = tap_get_tap_dev(q);
1094 		if (!tap) {
1095 			rtnl_unlock();
1096 			return -ENOLINK;
1097 		}
1098 		ret = 0;
1099 		u = tap->dev->type;
1100 		if (copy_to_user(&ifr->ifr_name, tap->dev->name, IFNAMSIZ) ||
1101 		    copy_to_user(&ifr->ifr_hwaddr.sa_data, tap->dev->dev_addr, ETH_ALEN) ||
1102 		    put_user(u, &ifr->ifr_hwaddr.sa_family))
1103 			ret = -EFAULT;
1104 		tap_put_tap_dev(tap);
1105 		rtnl_unlock();
1106 		return ret;
1107 
1108 	case SIOCSIFHWADDR:
1109 		if (copy_from_user(&sa, &ifr->ifr_hwaddr, sizeof(sa)))
1110 			return -EFAULT;
1111 		rtnl_lock();
1112 		tap = tap_get_tap_dev(q);
1113 		if (!tap) {
1114 			rtnl_unlock();
1115 			return -ENOLINK;
1116 		}
1117 		ret = dev_set_mac_address(tap->dev, &sa, NULL);
1118 		tap_put_tap_dev(tap);
1119 		rtnl_unlock();
1120 		return ret;
1121 
1122 	default:
1123 		return -EINVAL;
1124 	}
1125 }
1126 
1127 #ifdef CONFIG_COMPAT
1128 static long tap_compat_ioctl(struct file *file, unsigned int cmd,
1129 			     unsigned long arg)
1130 {
1131 	return tap_ioctl(file, cmd, (unsigned long)compat_ptr(arg));
1132 }
1133 #endif
1134 
1135 static const struct file_operations tap_fops = {
1136 	.owner		= THIS_MODULE,
1137 	.open		= tap_open,
1138 	.release	= tap_release,
1139 	.read_iter	= tap_read_iter,
1140 	.write_iter	= tap_write_iter,
1141 	.poll		= tap_poll,
1142 	.llseek		= no_llseek,
1143 	.unlocked_ioctl	= tap_ioctl,
1144 #ifdef CONFIG_COMPAT
1145 	.compat_ioctl	= tap_compat_ioctl,
1146 #endif
1147 };
1148 
1149 static int tap_get_user_xdp(struct tap_queue *q, struct xdp_buff *xdp)
1150 {
1151 	struct tun_xdp_hdr *hdr = xdp->data_hard_start;
1152 	struct virtio_net_hdr *gso = &hdr->gso;
1153 	int buflen = hdr->buflen;
1154 	int vnet_hdr_len = 0;
1155 	struct tap_dev *tap;
1156 	struct sk_buff *skb;
1157 	int err, depth;
1158 
1159 	if (q->flags & IFF_VNET_HDR)
1160 		vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
1161 
1162 	skb = build_skb(xdp->data_hard_start, buflen);
1163 	if (!skb) {
1164 		err = -ENOMEM;
1165 		goto err;
1166 	}
1167 
1168 	skb_reserve(skb, xdp->data - xdp->data_hard_start);
1169 	skb_put(skb, xdp->data_end - xdp->data);
1170 
1171 	skb_set_network_header(skb, ETH_HLEN);
1172 	skb_reset_mac_header(skb);
1173 	skb->protocol = eth_hdr(skb)->h_proto;
1174 
1175 	if (vnet_hdr_len) {
1176 		err = virtio_net_hdr_to_skb(skb, gso, tap_is_little_endian(q));
1177 		if (err)
1178 			goto err_kfree;
1179 	}
1180 
1181 	/* Move network header to the right position for VLAN tagged packets */
1182 	if ((skb->protocol == htons(ETH_P_8021Q) ||
1183 	     skb->protocol == htons(ETH_P_8021AD)) &&
1184 	    __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
1185 		skb_set_network_header(skb, depth);
1186 
1187 	rcu_read_lock();
1188 	tap = rcu_dereference(q->tap);
1189 	if (tap) {
1190 		skb->dev = tap->dev;
1191 		skb_probe_transport_header(skb);
1192 		dev_queue_xmit(skb);
1193 	} else {
1194 		kfree_skb(skb);
1195 	}
1196 	rcu_read_unlock();
1197 
1198 	return 0;
1199 
1200 err_kfree:
1201 	kfree_skb(skb);
1202 err:
1203 	rcu_read_lock();
1204 		tap = rcu_dereference(q->tap);
1205 	if (tap && tap->count_tx_dropped)
1206 		tap->count_tx_dropped(tap);
1207 	rcu_read_unlock();
1208 	return err;
1209 }
1210 
1211 static int tap_sendmsg(struct socket *sock, struct msghdr *m,
1212 		       size_t total_len)
1213 {
1214 	struct tap_queue *q = container_of(sock, struct tap_queue, sock);
1215 	struct tun_msg_ctl *ctl = m->msg_control;
1216 	struct xdp_buff *xdp;
1217 	int i;
1218 
1219 	if (ctl && (ctl->type == TUN_MSG_PTR)) {
1220 		for (i = 0; i < ctl->num; i++) {
1221 			xdp = &((struct xdp_buff *)ctl->ptr)[i];
1222 			tap_get_user_xdp(q, xdp);
1223 		}
1224 		return 0;
1225 	}
1226 
1227 	return tap_get_user(q, ctl ? ctl->ptr : NULL, &m->msg_iter,
1228 			    m->msg_flags & MSG_DONTWAIT);
1229 }
1230 
1231 static int tap_recvmsg(struct socket *sock, struct msghdr *m,
1232 		       size_t total_len, int flags)
1233 {
1234 	struct tap_queue *q = container_of(sock, struct tap_queue, sock);
1235 	struct sk_buff *skb = m->msg_control;
1236 	int ret;
1237 	if (flags & ~(MSG_DONTWAIT|MSG_TRUNC)) {
1238 		kfree_skb(skb);
1239 		return -EINVAL;
1240 	}
1241 	ret = tap_do_read(q, &m->msg_iter, flags & MSG_DONTWAIT, skb);
1242 	if (ret > total_len) {
1243 		m->msg_flags |= MSG_TRUNC;
1244 		ret = flags & MSG_TRUNC ? ret : total_len;
1245 	}
1246 	return ret;
1247 }
1248 
1249 static int tap_peek_len(struct socket *sock)
1250 {
1251 	struct tap_queue *q = container_of(sock, struct tap_queue,
1252 					       sock);
1253 	return PTR_RING_PEEK_CALL(&q->ring, __skb_array_len_with_tag);
1254 }
1255 
1256 /* Ops structure to mimic raw sockets with tun */
1257 static const struct proto_ops tap_socket_ops = {
1258 	.sendmsg = tap_sendmsg,
1259 	.recvmsg = tap_recvmsg,
1260 	.peek_len = tap_peek_len,
1261 };
1262 
1263 /* Get an underlying socket object from tun file.  Returns error unless file is
1264  * attached to a device.  The returned object works like a packet socket, it
1265  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1266  * holding a reference to the file for as long as the socket is in use. */
1267 struct socket *tap_get_socket(struct file *file)
1268 {
1269 	struct tap_queue *q;
1270 	if (file->f_op != &tap_fops)
1271 		return ERR_PTR(-EINVAL);
1272 	q = file->private_data;
1273 	if (!q)
1274 		return ERR_PTR(-EBADFD);
1275 	return &q->sock;
1276 }
1277 EXPORT_SYMBOL_GPL(tap_get_socket);
1278 
1279 struct ptr_ring *tap_get_ptr_ring(struct file *file)
1280 {
1281 	struct tap_queue *q;
1282 
1283 	if (file->f_op != &tap_fops)
1284 		return ERR_PTR(-EINVAL);
1285 	q = file->private_data;
1286 	if (!q)
1287 		return ERR_PTR(-EBADFD);
1288 	return &q->ring;
1289 }
1290 EXPORT_SYMBOL_GPL(tap_get_ptr_ring);
1291 
1292 int tap_queue_resize(struct tap_dev *tap)
1293 {
1294 	struct net_device *dev = tap->dev;
1295 	struct tap_queue *q;
1296 	struct ptr_ring **rings;
1297 	int n = tap->numqueues;
1298 	int ret, i = 0;
1299 
1300 	rings = kmalloc_array(n, sizeof(*rings), GFP_KERNEL);
1301 	if (!rings)
1302 		return -ENOMEM;
1303 
1304 	list_for_each_entry(q, &tap->queue_list, next)
1305 		rings[i++] = &q->ring;
1306 
1307 	ret = ptr_ring_resize_multiple(rings, n,
1308 				       dev->tx_queue_len, GFP_KERNEL,
1309 				       __skb_array_destroy_skb);
1310 
1311 	kfree(rings);
1312 	return ret;
1313 }
1314 EXPORT_SYMBOL_GPL(tap_queue_resize);
1315 
1316 static int tap_list_add(dev_t major, const char *device_name)
1317 {
1318 	struct major_info *tap_major;
1319 
1320 	tap_major = kzalloc(sizeof(*tap_major), GFP_ATOMIC);
1321 	if (!tap_major)
1322 		return -ENOMEM;
1323 
1324 	tap_major->major = MAJOR(major);
1325 
1326 	idr_init(&tap_major->minor_idr);
1327 	spin_lock_init(&tap_major->minor_lock);
1328 
1329 	tap_major->device_name = device_name;
1330 
1331 	list_add_tail_rcu(&tap_major->next, &major_list);
1332 	return 0;
1333 }
1334 
1335 int tap_create_cdev(struct cdev *tap_cdev, dev_t *tap_major,
1336 		    const char *device_name, struct module *module)
1337 {
1338 	int err;
1339 
1340 	err = alloc_chrdev_region(tap_major, 0, TAP_NUM_DEVS, device_name);
1341 	if (err)
1342 		goto out1;
1343 
1344 	cdev_init(tap_cdev, &tap_fops);
1345 	tap_cdev->owner = module;
1346 	err = cdev_add(tap_cdev, *tap_major, TAP_NUM_DEVS);
1347 	if (err)
1348 		goto out2;
1349 
1350 	err =  tap_list_add(*tap_major, device_name);
1351 	if (err)
1352 		goto out3;
1353 
1354 	return 0;
1355 
1356 out3:
1357 	cdev_del(tap_cdev);
1358 out2:
1359 	unregister_chrdev_region(*tap_major, TAP_NUM_DEVS);
1360 out1:
1361 	return err;
1362 }
1363 EXPORT_SYMBOL_GPL(tap_create_cdev);
1364 
1365 void tap_destroy_cdev(dev_t major, struct cdev *tap_cdev)
1366 {
1367 	struct major_info *tap_major, *tmp;
1368 
1369 	cdev_del(tap_cdev);
1370 	unregister_chrdev_region(major, TAP_NUM_DEVS);
1371 	list_for_each_entry_safe(tap_major, tmp, &major_list, next) {
1372 		if (tap_major->major == MAJOR(major)) {
1373 			idr_destroy(&tap_major->minor_idr);
1374 			list_del_rcu(&tap_major->next);
1375 			kfree_rcu(tap_major, rcu);
1376 		}
1377 	}
1378 }
1379 EXPORT_SYMBOL_GPL(tap_destroy_cdev);
1380 
1381 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1382 MODULE_AUTHOR("Sainath Grandhi <sainath.grandhi@intel.com>");
1383 MODULE_LICENSE("GPL");
1384