GNU Linux-libre 5.19-rc6-gnu
[releases.git] / drivers / net / tap.c
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         enum skb_drop_reason drop_reason;
326
327         tap = tap_dev_get_rcu(dev);
328         if (!tap)
329                 return RX_HANDLER_PASS;
330
331         q = tap_get_queue(tap, skb);
332         if (!q)
333                 return RX_HANDLER_PASS;
334
335         skb_push(skb, ETH_HLEN);
336
337         /* Apply the forward feature mask so that we perform segmentation
338          * according to users wishes.  This only works if VNET_HDR is
339          * enabled.
340          */
341         if (q->flags & IFF_VNET_HDR)
342                 features |= tap->tap_features;
343         if (netif_needs_gso(skb, features)) {
344                 struct sk_buff *segs = __skb_gso_segment(skb, features, false);
345                 struct sk_buff *next;
346
347                 if (IS_ERR(segs)) {
348                         drop_reason = SKB_DROP_REASON_SKB_GSO_SEG;
349                         goto drop;
350                 }
351
352                 if (!segs) {
353                         if (ptr_ring_produce(&q->ring, skb)) {
354                                 drop_reason = SKB_DROP_REASON_FULL_RING;
355                                 goto drop;
356                         }
357                         goto wake_up;
358                 }
359
360                 consume_skb(skb);
361                 skb_list_walk_safe(segs, skb, next) {
362                         skb_mark_not_on_list(skb);
363                         if (ptr_ring_produce(&q->ring, skb)) {
364                                 drop_reason = SKB_DROP_REASON_FULL_RING;
365                                 kfree_skb_reason(skb, drop_reason);
366                                 kfree_skb_list_reason(next, drop_reason);
367                                 break;
368                         }
369                 }
370         } else {
371                 /* If we receive a partial checksum and the tap side
372                  * doesn't support checksum offload, compute the checksum.
373                  * Note: it doesn't matter which checksum feature to
374                  *        check, we either support them all or none.
375                  */
376                 if (skb->ip_summed == CHECKSUM_PARTIAL &&
377                     !(features & NETIF_F_CSUM_MASK) &&
378                     skb_checksum_help(skb)) {
379                         drop_reason = SKB_DROP_REASON_SKB_CSUM;
380                         goto drop;
381                 }
382                 if (ptr_ring_produce(&q->ring, skb)) {
383                         drop_reason = SKB_DROP_REASON_FULL_RING;
384                         goto drop;
385                 }
386         }
387
388 wake_up:
389         wake_up_interruptible_poll(sk_sleep(&q->sk), EPOLLIN | EPOLLRDNORM | EPOLLRDBAND);
390         return RX_HANDLER_CONSUMED;
391
392 drop:
393         /* Count errors/drops only here, thus don't care about args. */
394         if (tap->count_rx_dropped)
395                 tap->count_rx_dropped(tap);
396         kfree_skb_reason(skb, drop_reason);
397         return RX_HANDLER_CONSUMED;
398 }
399 EXPORT_SYMBOL_GPL(tap_handle_frame);
400
401 static struct major_info *tap_get_major(int major)
402 {
403         struct major_info *tap_major;
404
405         list_for_each_entry_rcu(tap_major, &major_list, next) {
406                 if (tap_major->major == major)
407                         return tap_major;
408         }
409
410         return NULL;
411 }
412
413 int tap_get_minor(dev_t major, struct tap_dev *tap)
414 {
415         int retval = -ENOMEM;
416         struct major_info *tap_major;
417
418         rcu_read_lock();
419         tap_major = tap_get_major(MAJOR(major));
420         if (!tap_major) {
421                 retval = -EINVAL;
422                 goto unlock;
423         }
424
425         spin_lock(&tap_major->minor_lock);
426         retval = idr_alloc(&tap_major->minor_idr, tap, 1, TAP_NUM_DEVS, GFP_ATOMIC);
427         if (retval >= 0) {
428                 tap->minor = retval;
429         } else if (retval == -ENOSPC) {
430                 netdev_err(tap->dev, "Too many tap devices\n");
431                 retval = -EINVAL;
432         }
433         spin_unlock(&tap_major->minor_lock);
434
435 unlock:
436         rcu_read_unlock();
437         return retval < 0 ? retval : 0;
438 }
439 EXPORT_SYMBOL_GPL(tap_get_minor);
440
441 void tap_free_minor(dev_t major, struct tap_dev *tap)
442 {
443         struct major_info *tap_major;
444
445         rcu_read_lock();
446         tap_major = tap_get_major(MAJOR(major));
447         if (!tap_major) {
448                 goto unlock;
449         }
450
451         spin_lock(&tap_major->minor_lock);
452         if (tap->minor) {
453                 idr_remove(&tap_major->minor_idr, tap->minor);
454                 tap->minor = 0;
455         }
456         spin_unlock(&tap_major->minor_lock);
457
458 unlock:
459         rcu_read_unlock();
460 }
461 EXPORT_SYMBOL_GPL(tap_free_minor);
462
463 static struct tap_dev *dev_get_by_tap_file(int major, int minor)
464 {
465         struct net_device *dev = NULL;
466         struct tap_dev *tap;
467         struct major_info *tap_major;
468
469         rcu_read_lock();
470         tap_major = tap_get_major(major);
471         if (!tap_major) {
472                 tap = NULL;
473                 goto unlock;
474         }
475
476         spin_lock(&tap_major->minor_lock);
477         tap = idr_find(&tap_major->minor_idr, minor);
478         if (tap) {
479                 dev = tap->dev;
480                 dev_hold(dev);
481         }
482         spin_unlock(&tap_major->minor_lock);
483
484 unlock:
485         rcu_read_unlock();
486         return tap;
487 }
488
489 static void tap_sock_write_space(struct sock *sk)
490 {
491         wait_queue_head_t *wqueue;
492
493         if (!sock_writeable(sk) ||
494             !test_and_clear_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags))
495                 return;
496
497         wqueue = sk_sleep(sk);
498         if (wqueue && waitqueue_active(wqueue))
499                 wake_up_interruptible_poll(wqueue, EPOLLOUT | EPOLLWRNORM | EPOLLWRBAND);
500 }
501
502 static void tap_sock_destruct(struct sock *sk)
503 {
504         struct tap_queue *q = container_of(sk, struct tap_queue, sk);
505
506         ptr_ring_cleanup(&q->ring, __skb_array_destroy_skb);
507 }
508
509 static int tap_open(struct inode *inode, struct file *file)
510 {
511         struct net *net = current->nsproxy->net_ns;
512         struct tap_dev *tap;
513         struct tap_queue *q;
514         int err = -ENODEV;
515
516         rtnl_lock();
517         tap = dev_get_by_tap_file(imajor(inode), iminor(inode));
518         if (!tap)
519                 goto err;
520
521         err = -ENOMEM;
522         q = (struct tap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
523                                              &tap_proto, 0);
524         if (!q)
525                 goto err;
526         if (ptr_ring_init(&q->ring, tap->dev->tx_queue_len, GFP_KERNEL)) {
527                 sk_free(&q->sk);
528                 goto err;
529         }
530
531         init_waitqueue_head(&q->sock.wq.wait);
532         q->sock.type = SOCK_RAW;
533         q->sock.state = SS_CONNECTED;
534         q->sock.file = file;
535         q->sock.ops = &tap_socket_ops;
536         sock_init_data(&q->sock, &q->sk);
537         q->sk.sk_write_space = tap_sock_write_space;
538         q->sk.sk_destruct = tap_sock_destruct;
539         q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
540         q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
541
542         /*
543          * so far only KVM virtio_net uses tap, enable zero copy between
544          * guest kernel and host kernel when lower device supports zerocopy
545          *
546          * The macvlan supports zerocopy iff the lower device supports zero
547          * copy so we don't have to look at the lower device directly.
548          */
549         if ((tap->dev->features & NETIF_F_HIGHDMA) && (tap->dev->features & NETIF_F_SG))
550                 sock_set_flag(&q->sk, SOCK_ZEROCOPY);
551
552         err = tap_set_queue(tap, file, q);
553         if (err) {
554                 /* tap_sock_destruct() will take care of freeing ptr_ring */
555                 goto err_put;
556         }
557
558         dev_put(tap->dev);
559
560         rtnl_unlock();
561         return err;
562
563 err_put:
564         sock_put(&q->sk);
565 err:
566         if (tap)
567                 dev_put(tap->dev);
568
569         rtnl_unlock();
570         return err;
571 }
572
573 static int tap_release(struct inode *inode, struct file *file)
574 {
575         struct tap_queue *q = file->private_data;
576         tap_put_queue(q);
577         return 0;
578 }
579
580 static __poll_t tap_poll(struct file *file, poll_table *wait)
581 {
582         struct tap_queue *q = file->private_data;
583         __poll_t mask = EPOLLERR;
584
585         if (!q)
586                 goto out;
587
588         mask = 0;
589         poll_wait(file, &q->sock.wq.wait, wait);
590
591         if (!ptr_ring_empty(&q->ring))
592                 mask |= EPOLLIN | EPOLLRDNORM;
593
594         if (sock_writeable(&q->sk) ||
595             (!test_and_set_bit(SOCKWQ_ASYNC_NOSPACE, &q->sock.flags) &&
596              sock_writeable(&q->sk)))
597                 mask |= EPOLLOUT | EPOLLWRNORM;
598
599 out:
600         return mask;
601 }
602
603 static inline struct sk_buff *tap_alloc_skb(struct sock *sk, size_t prepad,
604                                             size_t len, size_t linear,
605                                                 int noblock, int *err)
606 {
607         struct sk_buff *skb;
608
609         /* Under a page?  Don't bother with paged skb. */
610         if (prepad + len < PAGE_SIZE || !linear)
611                 linear = len;
612
613         skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
614                                    err, 0);
615         if (!skb)
616                 return NULL;
617
618         skb_reserve(skb, prepad);
619         skb_put(skb, linear);
620         skb->data_len = len - linear;
621         skb->len += len - linear;
622
623         return skb;
624 }
625
626 /* Neighbour code has some assumptions on HH_DATA_MOD alignment */
627 #define TAP_RESERVE HH_DATA_OFF(ETH_HLEN)
628
629 /* Get packet from user space buffer */
630 static ssize_t tap_get_user(struct tap_queue *q, void *msg_control,
631                             struct iov_iter *from, int noblock)
632 {
633         int good_linear = SKB_MAX_HEAD(TAP_RESERVE);
634         struct sk_buff *skb;
635         struct tap_dev *tap;
636         unsigned long total_len = iov_iter_count(from);
637         unsigned long len = total_len;
638         int err;
639         struct virtio_net_hdr vnet_hdr = { 0 };
640         int vnet_hdr_len = 0;
641         int copylen = 0;
642         int depth;
643         bool zerocopy = false;
644         size_t linear;
645         enum skb_drop_reason drop_reason;
646
647         if (q->flags & IFF_VNET_HDR) {
648                 vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
649
650                 err = -EINVAL;
651                 if (len < vnet_hdr_len)
652                         goto err;
653                 len -= vnet_hdr_len;
654
655                 err = -EFAULT;
656                 if (!copy_from_iter_full(&vnet_hdr, sizeof(vnet_hdr), from))
657                         goto err;
658                 iov_iter_advance(from, vnet_hdr_len - sizeof(vnet_hdr));
659                 if ((vnet_hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
660                      tap16_to_cpu(q, vnet_hdr.csum_start) +
661                      tap16_to_cpu(q, vnet_hdr.csum_offset) + 2 >
662                              tap16_to_cpu(q, vnet_hdr.hdr_len))
663                         vnet_hdr.hdr_len = cpu_to_tap16(q,
664                                  tap16_to_cpu(q, vnet_hdr.csum_start) +
665                                  tap16_to_cpu(q, vnet_hdr.csum_offset) + 2);
666                 err = -EINVAL;
667                 if (tap16_to_cpu(q, vnet_hdr.hdr_len) > len)
668                         goto err;
669         }
670
671         err = -EINVAL;
672         if (unlikely(len < ETH_HLEN))
673                 goto err;
674
675         if (msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
676                 struct iov_iter i;
677
678                 copylen = vnet_hdr.hdr_len ?
679                         tap16_to_cpu(q, vnet_hdr.hdr_len) : GOODCOPY_LEN;
680                 if (copylen > good_linear)
681                         copylen = good_linear;
682                 else if (copylen < ETH_HLEN)
683                         copylen = ETH_HLEN;
684                 linear = copylen;
685                 i = *from;
686                 iov_iter_advance(&i, copylen);
687                 if (iov_iter_npages(&i, INT_MAX) <= MAX_SKB_FRAGS)
688                         zerocopy = true;
689         }
690
691         if (!zerocopy) {
692                 copylen = len;
693                 linear = tap16_to_cpu(q, vnet_hdr.hdr_len);
694                 if (linear > good_linear)
695                         linear = good_linear;
696                 else if (linear < ETH_HLEN)
697                         linear = ETH_HLEN;
698         }
699
700         skb = tap_alloc_skb(&q->sk, TAP_RESERVE, copylen,
701                             linear, noblock, &err);
702         if (!skb)
703                 goto err;
704
705         if (zerocopy)
706                 err = zerocopy_sg_from_iter(skb, from);
707         else
708                 err = skb_copy_datagram_from_iter(skb, 0, from, len);
709
710         if (err) {
711                 drop_reason = SKB_DROP_REASON_SKB_UCOPY_FAULT;
712                 goto err_kfree;
713         }
714
715         skb_set_network_header(skb, ETH_HLEN);
716         skb_reset_mac_header(skb);
717         skb->protocol = eth_hdr(skb)->h_proto;
718
719         if (vnet_hdr_len) {
720                 err = virtio_net_hdr_to_skb(skb, &vnet_hdr,
721                                             tap_is_little_endian(q));
722                 if (err) {
723                         drop_reason = SKB_DROP_REASON_DEV_HDR;
724                         goto err_kfree;
725                 }
726         }
727
728         skb_probe_transport_header(skb);
729
730         /* Move network header to the right position for VLAN tagged packets */
731         if (eth_type_vlan(skb->protocol) &&
732             __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
733                 skb_set_network_header(skb, depth);
734
735         rcu_read_lock();
736         tap = rcu_dereference(q->tap);
737         /* copy skb_ubuf_info for callback when skb has no error */
738         if (zerocopy) {
739                 skb_zcopy_init(skb, msg_control);
740         } else if (msg_control) {
741                 struct ubuf_info *uarg = msg_control;
742                 uarg->callback(NULL, uarg, false);
743         }
744
745         if (tap) {
746                 skb->dev = tap->dev;
747                 dev_queue_xmit(skb);
748         } else {
749                 kfree_skb(skb);
750         }
751         rcu_read_unlock();
752
753         return total_len;
754
755 err_kfree:
756         kfree_skb_reason(skb, drop_reason);
757
758 err:
759         rcu_read_lock();
760         tap = rcu_dereference(q->tap);
761         if (tap && tap->count_tx_dropped)
762                 tap->count_tx_dropped(tap);
763         rcu_read_unlock();
764
765         return err;
766 }
767
768 static ssize_t tap_write_iter(struct kiocb *iocb, struct iov_iter *from)
769 {
770         struct file *file = iocb->ki_filp;
771         struct tap_queue *q = file->private_data;
772
773         return tap_get_user(q, NULL, from, file->f_flags & O_NONBLOCK);
774 }
775
776 /* Put packet to the user space buffer */
777 static ssize_t tap_put_user(struct tap_queue *q,
778                             const struct sk_buff *skb,
779                             struct iov_iter *iter)
780 {
781         int ret;
782         int vnet_hdr_len = 0;
783         int vlan_offset = 0;
784         int total;
785
786         if (q->flags & IFF_VNET_HDR) {
787                 int vlan_hlen = skb_vlan_tag_present(skb) ? VLAN_HLEN : 0;
788                 struct virtio_net_hdr vnet_hdr;
789
790                 vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
791                 if (iov_iter_count(iter) < vnet_hdr_len)
792                         return -EINVAL;
793
794                 if (virtio_net_hdr_from_skb(skb, &vnet_hdr,
795                                             tap_is_little_endian(q), true,
796                                             vlan_hlen))
797                         BUG();
798
799                 if (copy_to_iter(&vnet_hdr, sizeof(vnet_hdr), iter) !=
800                     sizeof(vnet_hdr))
801                         return -EFAULT;
802
803                 iov_iter_advance(iter, vnet_hdr_len - sizeof(vnet_hdr));
804         }
805         total = vnet_hdr_len;
806         total += skb->len;
807
808         if (skb_vlan_tag_present(skb)) {
809                 struct {
810                         __be16 h_vlan_proto;
811                         __be16 h_vlan_TCI;
812                 } veth;
813                 veth.h_vlan_proto = skb->vlan_proto;
814                 veth.h_vlan_TCI = htons(skb_vlan_tag_get(skb));
815
816                 vlan_offset = offsetof(struct vlan_ethhdr, h_vlan_proto);
817                 total += VLAN_HLEN;
818
819                 ret = skb_copy_datagram_iter(skb, 0, iter, vlan_offset);
820                 if (ret || !iov_iter_count(iter))
821                         goto done;
822
823                 ret = copy_to_iter(&veth, sizeof(veth), iter);
824                 if (ret != sizeof(veth) || !iov_iter_count(iter))
825                         goto done;
826         }
827
828         ret = skb_copy_datagram_iter(skb, vlan_offset, iter,
829                                      skb->len - vlan_offset);
830
831 done:
832         return ret ? ret : total;
833 }
834
835 static ssize_t tap_do_read(struct tap_queue *q,
836                            struct iov_iter *to,
837                            int noblock, struct sk_buff *skb)
838 {
839         DEFINE_WAIT(wait);
840         ssize_t ret = 0;
841
842         if (!iov_iter_count(to)) {
843                 kfree_skb(skb);
844                 return 0;
845         }
846
847         if (skb)
848                 goto put;
849
850         while (1) {
851                 if (!noblock)
852                         prepare_to_wait(sk_sleep(&q->sk), &wait,
853                                         TASK_INTERRUPTIBLE);
854
855                 /* Read frames from the queue */
856                 skb = ptr_ring_consume(&q->ring);
857                 if (skb)
858                         break;
859                 if (noblock) {
860                         ret = -EAGAIN;
861                         break;
862                 }
863                 if (signal_pending(current)) {
864                         ret = -ERESTARTSYS;
865                         break;
866                 }
867                 /* Nothing to read, let's sleep */
868                 schedule();
869         }
870         if (!noblock)
871                 finish_wait(sk_sleep(&q->sk), &wait);
872
873 put:
874         if (skb) {
875                 ret = tap_put_user(q, skb, to);
876                 if (unlikely(ret < 0))
877                         kfree_skb(skb);
878                 else
879                         consume_skb(skb);
880         }
881         return ret;
882 }
883
884 static ssize_t tap_read_iter(struct kiocb *iocb, struct iov_iter *to)
885 {
886         struct file *file = iocb->ki_filp;
887         struct tap_queue *q = file->private_data;
888         ssize_t len = iov_iter_count(to), ret;
889
890         ret = tap_do_read(q, to, file->f_flags & O_NONBLOCK, NULL);
891         ret = min_t(ssize_t, ret, len);
892         if (ret > 0)
893                 iocb->ki_pos = ret;
894         return ret;
895 }
896
897 static struct tap_dev *tap_get_tap_dev(struct tap_queue *q)
898 {
899         struct tap_dev *tap;
900
901         ASSERT_RTNL();
902         tap = rtnl_dereference(q->tap);
903         if (tap)
904                 dev_hold(tap->dev);
905
906         return tap;
907 }
908
909 static void tap_put_tap_dev(struct tap_dev *tap)
910 {
911         dev_put(tap->dev);
912 }
913
914 static int tap_ioctl_set_queue(struct file *file, unsigned int flags)
915 {
916         struct tap_queue *q = file->private_data;
917         struct tap_dev *tap;
918         int ret;
919
920         tap = tap_get_tap_dev(q);
921         if (!tap)
922                 return -EINVAL;
923
924         if (flags & IFF_ATTACH_QUEUE)
925                 ret = tap_enable_queue(tap, file, q);
926         else if (flags & IFF_DETACH_QUEUE)
927                 ret = tap_disable_queue(q);
928         else
929                 ret = -EINVAL;
930
931         tap_put_tap_dev(tap);
932         return ret;
933 }
934
935 static int set_offload(struct tap_queue *q, unsigned long arg)
936 {
937         struct tap_dev *tap;
938         netdev_features_t features;
939         netdev_features_t feature_mask = 0;
940
941         tap = rtnl_dereference(q->tap);
942         if (!tap)
943                 return -ENOLINK;
944
945         features = tap->dev->features;
946
947         if (arg & TUN_F_CSUM) {
948                 feature_mask = NETIF_F_HW_CSUM;
949
950                 if (arg & (TUN_F_TSO4 | TUN_F_TSO6)) {
951                         if (arg & TUN_F_TSO_ECN)
952                                 feature_mask |= NETIF_F_TSO_ECN;
953                         if (arg & TUN_F_TSO4)
954                                 feature_mask |= NETIF_F_TSO;
955                         if (arg & TUN_F_TSO6)
956                                 feature_mask |= NETIF_F_TSO6;
957                 }
958         }
959
960         /* tun/tap driver inverts the usage for TSO offloads, where
961          * setting the TSO bit means that the userspace wants to
962          * accept TSO frames and turning it off means that user space
963          * does not support TSO.
964          * For tap, we have to invert it to mean the same thing.
965          * When user space turns off TSO, we turn off GSO/LRO so that
966          * user-space will not receive TSO frames.
967          */
968         if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6))
969                 features |= RX_OFFLOADS;
970         else
971                 features &= ~RX_OFFLOADS;
972
973         /* tap_features are the same as features on tun/tap and
974          * reflect user expectations.
975          */
976         tap->tap_features = feature_mask;
977         if (tap->update_features)
978                 tap->update_features(tap, features);
979
980         return 0;
981 }
982
983 /*
984  * provide compatibility with generic tun/tap interface
985  */
986 static long tap_ioctl(struct file *file, unsigned int cmd,
987                       unsigned long arg)
988 {
989         struct tap_queue *q = file->private_data;
990         struct tap_dev *tap;
991         void __user *argp = (void __user *)arg;
992         struct ifreq __user *ifr = argp;
993         unsigned int __user *up = argp;
994         unsigned short u;
995         int __user *sp = argp;
996         struct sockaddr sa;
997         int s;
998         int ret;
999
1000         switch (cmd) {
1001         case TUNSETIFF:
1002                 /* ignore the name, just look at flags */
1003                 if (get_user(u, &ifr->ifr_flags))
1004                         return -EFAULT;
1005
1006                 ret = 0;
1007                 if ((u & ~TAP_IFFEATURES) != (IFF_NO_PI | IFF_TAP))
1008                         ret = -EINVAL;
1009                 else
1010                         q->flags = (q->flags & ~TAP_IFFEATURES) | u;
1011
1012                 return ret;
1013
1014         case TUNGETIFF:
1015                 rtnl_lock();
1016                 tap = tap_get_tap_dev(q);
1017                 if (!tap) {
1018                         rtnl_unlock();
1019                         return -ENOLINK;
1020                 }
1021
1022                 ret = 0;
1023                 u = q->flags;
1024                 if (copy_to_user(&ifr->ifr_name, tap->dev->name, IFNAMSIZ) ||
1025                     put_user(u, &ifr->ifr_flags))
1026                         ret = -EFAULT;
1027                 tap_put_tap_dev(tap);
1028                 rtnl_unlock();
1029                 return ret;
1030
1031         case TUNSETQUEUE:
1032                 if (get_user(u, &ifr->ifr_flags))
1033                         return -EFAULT;
1034                 rtnl_lock();
1035                 ret = tap_ioctl_set_queue(file, u);
1036                 rtnl_unlock();
1037                 return ret;
1038
1039         case TUNGETFEATURES:
1040                 if (put_user(IFF_TAP | IFF_NO_PI | TAP_IFFEATURES, up))
1041                         return -EFAULT;
1042                 return 0;
1043
1044         case TUNSETSNDBUF:
1045                 if (get_user(s, sp))
1046                         return -EFAULT;
1047                 if (s <= 0)
1048                         return -EINVAL;
1049
1050                 q->sk.sk_sndbuf = s;
1051                 return 0;
1052
1053         case TUNGETVNETHDRSZ:
1054                 s = q->vnet_hdr_sz;
1055                 if (put_user(s, sp))
1056                         return -EFAULT;
1057                 return 0;
1058
1059         case TUNSETVNETHDRSZ:
1060                 if (get_user(s, sp))
1061                         return -EFAULT;
1062                 if (s < (int)sizeof(struct virtio_net_hdr))
1063                         return -EINVAL;
1064
1065                 q->vnet_hdr_sz = s;
1066                 return 0;
1067
1068         case TUNGETVNETLE:
1069                 s = !!(q->flags & TAP_VNET_LE);
1070                 if (put_user(s, sp))
1071                         return -EFAULT;
1072                 return 0;
1073
1074         case TUNSETVNETLE:
1075                 if (get_user(s, sp))
1076                         return -EFAULT;
1077                 if (s)
1078                         q->flags |= TAP_VNET_LE;
1079                 else
1080                         q->flags &= ~TAP_VNET_LE;
1081                 return 0;
1082
1083         case TUNGETVNETBE:
1084                 return tap_get_vnet_be(q, sp);
1085
1086         case TUNSETVNETBE:
1087                 return tap_set_vnet_be(q, sp);
1088
1089         case TUNSETOFFLOAD:
1090                 /* let the user check for future flags */
1091                 if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1092                             TUN_F_TSO_ECN | TUN_F_UFO))
1093                         return -EINVAL;
1094
1095                 rtnl_lock();
1096                 ret = set_offload(q, arg);
1097                 rtnl_unlock();
1098                 return ret;
1099
1100         case SIOCGIFHWADDR:
1101                 rtnl_lock();
1102                 tap = tap_get_tap_dev(q);
1103                 if (!tap) {
1104                         rtnl_unlock();
1105                         return -ENOLINK;
1106                 }
1107                 ret = 0;
1108                 dev_get_mac_address(&sa, dev_net(tap->dev), tap->dev->name);
1109                 if (copy_to_user(&ifr->ifr_name, tap->dev->name, IFNAMSIZ) ||
1110                     copy_to_user(&ifr->ifr_hwaddr, &sa, sizeof(sa)))
1111                         ret = -EFAULT;
1112                 tap_put_tap_dev(tap);
1113                 rtnl_unlock();
1114                 return ret;
1115
1116         case SIOCSIFHWADDR:
1117                 if (copy_from_user(&sa, &ifr->ifr_hwaddr, sizeof(sa)))
1118                         return -EFAULT;
1119                 rtnl_lock();
1120                 tap = tap_get_tap_dev(q);
1121                 if (!tap) {
1122                         rtnl_unlock();
1123                         return -ENOLINK;
1124                 }
1125                 ret = dev_set_mac_address_user(tap->dev, &sa, NULL);
1126                 tap_put_tap_dev(tap);
1127                 rtnl_unlock();
1128                 return ret;
1129
1130         default:
1131                 return -EINVAL;
1132         }
1133 }
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         .compat_ioctl   = compat_ptr_ioctl,
1145 };
1146
1147 static int tap_get_user_xdp(struct tap_queue *q, struct xdp_buff *xdp)
1148 {
1149         struct tun_xdp_hdr *hdr = xdp->data_hard_start;
1150         struct virtio_net_hdr *gso = &hdr->gso;
1151         int buflen = hdr->buflen;
1152         int vnet_hdr_len = 0;
1153         struct tap_dev *tap;
1154         struct sk_buff *skb;
1155         int err, depth;
1156
1157         if (q->flags & IFF_VNET_HDR)
1158                 vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
1159
1160         skb = build_skb(xdp->data_hard_start, buflen);
1161         if (!skb) {
1162                 err = -ENOMEM;
1163                 goto err;
1164         }
1165
1166         skb_reserve(skb, xdp->data - xdp->data_hard_start);
1167         skb_put(skb, xdp->data_end - xdp->data);
1168
1169         skb_set_network_header(skb, ETH_HLEN);
1170         skb_reset_mac_header(skb);
1171         skb->protocol = eth_hdr(skb)->h_proto;
1172
1173         if (vnet_hdr_len) {
1174                 err = virtio_net_hdr_to_skb(skb, gso, tap_is_little_endian(q));
1175                 if (err)
1176                         goto err_kfree;
1177         }
1178
1179         /* Move network header to the right position for VLAN tagged packets */
1180         if (eth_type_vlan(skb->protocol) &&
1181             __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
1182                 skb_set_network_header(skb, depth);
1183
1184         rcu_read_lock();
1185         tap = rcu_dereference(q->tap);
1186         if (tap) {
1187                 skb->dev = tap->dev;
1188                 skb_probe_transport_header(skb);
1189                 dev_queue_xmit(skb);
1190         } else {
1191                 kfree_skb(skb);
1192         }
1193         rcu_read_unlock();
1194
1195         return 0;
1196
1197 err_kfree:
1198         kfree_skb(skb);
1199 err:
1200         rcu_read_lock();
1201         tap = rcu_dereference(q->tap);
1202         if (tap && tap->count_tx_dropped)
1203                 tap->count_tx_dropped(tap);
1204         rcu_read_unlock();
1205         return err;
1206 }
1207
1208 static int tap_sendmsg(struct socket *sock, struct msghdr *m,
1209                        size_t total_len)
1210 {
1211         struct tap_queue *q = container_of(sock, struct tap_queue, sock);
1212         struct tun_msg_ctl *ctl = m->msg_control;
1213         struct xdp_buff *xdp;
1214         int i;
1215
1216         if (m->msg_controllen == sizeof(struct tun_msg_ctl) &&
1217             ctl && ctl->type == TUN_MSG_PTR) {
1218                 for (i = 0; i < ctl->num; i++) {
1219                         xdp = &((struct xdp_buff *)ctl->ptr)[i];
1220                         tap_get_user_xdp(q, xdp);
1221                 }
1222                 return 0;
1223         }
1224
1225         return tap_get_user(q, ctl ? ctl->ptr : NULL, &m->msg_iter,
1226                             m->msg_flags & MSG_DONTWAIT);
1227 }
1228
1229 static int tap_recvmsg(struct socket *sock, struct msghdr *m,
1230                        size_t total_len, int flags)
1231 {
1232         struct tap_queue *q = container_of(sock, struct tap_queue, sock);
1233         struct sk_buff *skb = m->msg_control;
1234         int ret;
1235         if (flags & ~(MSG_DONTWAIT|MSG_TRUNC)) {
1236                 kfree_skb(skb);
1237                 return -EINVAL;
1238         }
1239         ret = tap_do_read(q, &m->msg_iter, flags & MSG_DONTWAIT, skb);
1240         if (ret > total_len) {
1241                 m->msg_flags |= MSG_TRUNC;
1242                 ret = flags & MSG_TRUNC ? ret : total_len;
1243         }
1244         return ret;
1245 }
1246
1247 static int tap_peek_len(struct socket *sock)
1248 {
1249         struct tap_queue *q = container_of(sock, struct tap_queue,
1250                                                sock);
1251         return PTR_RING_PEEK_CALL(&q->ring, __skb_array_len_with_tag);
1252 }
1253
1254 /* Ops structure to mimic raw sockets with tun */
1255 static const struct proto_ops tap_socket_ops = {
1256         .sendmsg = tap_sendmsg,
1257         .recvmsg = tap_recvmsg,
1258         .peek_len = tap_peek_len,
1259 };
1260
1261 /* Get an underlying socket object from tun file.  Returns error unless file is
1262  * attached to a device.  The returned object works like a packet socket, it
1263  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1264  * holding a reference to the file for as long as the socket is in use. */
1265 struct socket *tap_get_socket(struct file *file)
1266 {
1267         struct tap_queue *q;
1268         if (file->f_op != &tap_fops)
1269                 return ERR_PTR(-EINVAL);
1270         q = file->private_data;
1271         if (!q)
1272                 return ERR_PTR(-EBADFD);
1273         return &q->sock;
1274 }
1275 EXPORT_SYMBOL_GPL(tap_get_socket);
1276
1277 struct ptr_ring *tap_get_ptr_ring(struct file *file)
1278 {
1279         struct tap_queue *q;
1280
1281         if (file->f_op != &tap_fops)
1282                 return ERR_PTR(-EINVAL);
1283         q = file->private_data;
1284         if (!q)
1285                 return ERR_PTR(-EBADFD);
1286         return &q->ring;
1287 }
1288 EXPORT_SYMBOL_GPL(tap_get_ptr_ring);
1289
1290 int tap_queue_resize(struct tap_dev *tap)
1291 {
1292         struct net_device *dev = tap->dev;
1293         struct tap_queue *q;
1294         struct ptr_ring **rings;
1295         int n = tap->numqueues;
1296         int ret, i = 0;
1297
1298         rings = kmalloc_array(n, sizeof(*rings), GFP_KERNEL);
1299         if (!rings)
1300                 return -ENOMEM;
1301
1302         list_for_each_entry(q, &tap->queue_list, next)
1303                 rings[i++] = &q->ring;
1304
1305         ret = ptr_ring_resize_multiple(rings, n,
1306                                        dev->tx_queue_len, GFP_KERNEL,
1307                                        __skb_array_destroy_skb);
1308
1309         kfree(rings);
1310         return ret;
1311 }
1312 EXPORT_SYMBOL_GPL(tap_queue_resize);
1313
1314 static int tap_list_add(dev_t major, const char *device_name)
1315 {
1316         struct major_info *tap_major;
1317
1318         tap_major = kzalloc(sizeof(*tap_major), GFP_ATOMIC);
1319         if (!tap_major)
1320                 return -ENOMEM;
1321
1322         tap_major->major = MAJOR(major);
1323
1324         idr_init(&tap_major->minor_idr);
1325         spin_lock_init(&tap_major->minor_lock);
1326
1327         tap_major->device_name = device_name;
1328
1329         list_add_tail_rcu(&tap_major->next, &major_list);
1330         return 0;
1331 }
1332
1333 int tap_create_cdev(struct cdev *tap_cdev, dev_t *tap_major,
1334                     const char *device_name, struct module *module)
1335 {
1336         int err;
1337
1338         err = alloc_chrdev_region(tap_major, 0, TAP_NUM_DEVS, device_name);
1339         if (err)
1340                 goto out1;
1341
1342         cdev_init(tap_cdev, &tap_fops);
1343         tap_cdev->owner = module;
1344         err = cdev_add(tap_cdev, *tap_major, TAP_NUM_DEVS);
1345         if (err)
1346                 goto out2;
1347
1348         err =  tap_list_add(*tap_major, device_name);
1349         if (err)
1350                 goto out3;
1351
1352         return 0;
1353
1354 out3:
1355         cdev_del(tap_cdev);
1356 out2:
1357         unregister_chrdev_region(*tap_major, TAP_NUM_DEVS);
1358 out1:
1359         return err;
1360 }
1361 EXPORT_SYMBOL_GPL(tap_create_cdev);
1362
1363 void tap_destroy_cdev(dev_t major, struct cdev *tap_cdev)
1364 {
1365         struct major_info *tap_major, *tmp;
1366
1367         cdev_del(tap_cdev);
1368         unregister_chrdev_region(major, TAP_NUM_DEVS);
1369         list_for_each_entry_safe(tap_major, tmp, &major_list, next) {
1370                 if (tap_major->major == MAJOR(major)) {
1371                         idr_destroy(&tap_major->minor_idr);
1372                         list_del_rcu(&tap_major->next);
1373                         kfree_rcu(tap_major, rcu);
1374                 }
1375         }
1376 }
1377 EXPORT_SYMBOL_GPL(tap_destroy_cdev);
1378
1379 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1380 MODULE_AUTHOR("Sainath Grandhi <sainath.grandhi@intel.com>");
1381 MODULE_LICENSE("GPL");