]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * TUN - Universal TUN/TAP device driver. | |
3 | * Copyright (C) 1999-2002 Maxim Krasnyansky <maxk@qualcomm.com> | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License as published by | |
7 | * the Free Software Foundation; either version 2 of the License, or | |
8 | * (at your option) any later version. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, | |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 | * GNU General Public License for more details. | |
14 | * | |
15 | * $Id: tun.c,v 1.15 2002/03/01 02:44:24 maxk Exp $ | |
16 | */ | |
17 | ||
18 | /* | |
19 | * Changes: | |
20 | * | |
ff4cc3ac MK |
21 | * Mike Kershaw <dragorn@kismetwireless.net> 2005/08/14 |
22 | * Add TUNSETLINK ioctl to set the link encapsulation | |
23 | * | |
1da177e4 | 24 | * Mark Smith <markzzzsmith@yahoo.com.au> |
344dc8ed | 25 | * Use eth_random_addr() for tap MAC address. |
1da177e4 LT |
26 | * |
27 | * Harald Roelle <harald.roelle@ifi.lmu.de> 2004/04/20 | |
28 | * Fixes in packet dropping, queue length setting and queue wakeup. | |
29 | * Increased default tx queue length. | |
30 | * Added ethtool API. | |
31 | * Minor cleanups | |
32 | * | |
33 | * Daniel Podlejski <underley@underley.eu.org> | |
34 | * Modifications for 2.3.99-pre5 kernel. | |
35 | */ | |
36 | ||
6b8a66ee JP |
37 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
38 | ||
1da177e4 LT |
39 | #define DRV_NAME "tun" |
40 | #define DRV_VERSION "1.6" | |
41 | #define DRV_DESCRIPTION "Universal TUN/TAP device driver" | |
42 | #define DRV_COPYRIGHT "(C) 1999-2004 Max Krasnyansky <maxk@qualcomm.com>" | |
43 | ||
1da177e4 LT |
44 | #include <linux/module.h> |
45 | #include <linux/errno.h> | |
46 | #include <linux/kernel.h> | |
174cd4b1 | 47 | #include <linux/sched/signal.h> |
1da177e4 LT |
48 | #include <linux/major.h> |
49 | #include <linux/slab.h> | |
50 | #include <linux/poll.h> | |
51 | #include <linux/fcntl.h> | |
52 | #include <linux/init.h> | |
53 | #include <linux/skbuff.h> | |
54 | #include <linux/netdevice.h> | |
55 | #include <linux/etherdevice.h> | |
56 | #include <linux/miscdevice.h> | |
57 | #include <linux/ethtool.h> | |
58 | #include <linux/rtnetlink.h> | |
50857e2a | 59 | #include <linux/compat.h> |
1da177e4 LT |
60 | #include <linux/if.h> |
61 | #include <linux/if_arp.h> | |
62 | #include <linux/if_ether.h> | |
63 | #include <linux/if_tun.h> | |
6680ec68 | 64 | #include <linux/if_vlan.h> |
1da177e4 | 65 | #include <linux/crc32.h> |
d647a591 | 66 | #include <linux/nsproxy.h> |
f43798c2 | 67 | #include <linux/virtio_net.h> |
99405162 | 68 | #include <linux/rcupdate.h> |
881d966b | 69 | #include <net/net_namespace.h> |
79d17604 | 70 | #include <net/netns/generic.h> |
f019a7a5 | 71 | #include <net/rtnetlink.h> |
33dccbb0 | 72 | #include <net/sock.h> |
93e14b6d | 73 | #include <linux/seq_file.h> |
e0b46d0e | 74 | #include <linux/uio.h> |
1576d986 | 75 | #include <linux/skb_array.h> |
1da177e4 | 76 | |
7c0f6ba6 | 77 | #include <linux/uaccess.h> |
1da177e4 | 78 | |
14daa021 RR |
79 | /* Uncomment to enable debugging */ |
80 | /* #define TUN_DEBUG 1 */ | |
81 | ||
1da177e4 LT |
82 | #ifdef TUN_DEBUG |
83 | static int debug; | |
14daa021 | 84 | |
6b8a66ee JP |
85 | #define tun_debug(level, tun, fmt, args...) \ |
86 | do { \ | |
87 | if (tun->debug) \ | |
88 | netdev_printk(level, tun->dev, fmt, ##args); \ | |
89 | } while (0) | |
90 | #define DBG1(level, fmt, args...) \ | |
91 | do { \ | |
92 | if (debug == 2) \ | |
93 | printk(level fmt, ##args); \ | |
94 | } while (0) | |
14daa021 | 95 | #else |
6b8a66ee JP |
96 | #define tun_debug(level, tun, fmt, args...) \ |
97 | do { \ | |
98 | if (0) \ | |
99 | netdev_printk(level, tun->dev, fmt, ##args); \ | |
100 | } while (0) | |
101 | #define DBG1(level, fmt, args...) \ | |
102 | do { \ | |
103 | if (0) \ | |
104 | printk(level fmt, ##args); \ | |
105 | } while (0) | |
14daa021 RR |
106 | #endif |
107 | ||
031f5e03 MT |
108 | /* TUN device flags */ |
109 | ||
110 | /* IFF_ATTACH_QUEUE is never stored in device flags, | |
111 | * overload it to mean fasync when stored there. | |
112 | */ | |
113 | #define TUN_FASYNC IFF_ATTACH_QUEUE | |
1cf8e410 MT |
114 | /* High bits in flags field are unused. */ |
115 | #define TUN_VNET_LE 0x80000000 | |
8b8e658b | 116 | #define TUN_VNET_BE 0x40000000 |
031f5e03 MT |
117 | |
118 | #define TUN_FEATURES (IFF_NO_PI | IFF_ONE_QUEUE | IFF_VNET_HDR | \ | |
1cf8e410 | 119 | IFF_MULTI_QUEUE) |
0690899b MT |
120 | #define GOODCOPY_LEN 128 |
121 | ||
f271b2cc MK |
122 | #define FLT_EXACT_COUNT 8 |
123 | struct tap_filter { | |
124 | unsigned int count; /* Number of addrs. Zero means disabled */ | |
125 | u32 mask[2]; /* Mask of the hashed addrs */ | |
126 | unsigned char addr[FLT_EXACT_COUNT][ETH_ALEN]; | |
127 | }; | |
128 | ||
baf71c5c PG |
129 | /* MAX_TAP_QUEUES 256 is chosen to allow rx/tx queues to be equal |
130 | * to max number of VCPUs in guest. */ | |
131 | #define MAX_TAP_QUEUES 256 | |
b8732fb7 | 132 | #define MAX_TAP_FLOWS 4096 |
c8d68e6b | 133 | |
96442e42 JW |
134 | #define TUN_FLOW_EXPIRE (3 * HZ) |
135 | ||
608b9977 PA |
136 | struct tun_pcpu_stats { |
137 | u64 rx_packets; | |
138 | u64 rx_bytes; | |
139 | u64 tx_packets; | |
140 | u64 tx_bytes; | |
141 | struct u64_stats_sync syncp; | |
142 | u32 rx_dropped; | |
143 | u32 tx_dropped; | |
144 | u32 rx_frame_errors; | |
145 | }; | |
146 | ||
54f968d6 | 147 | /* A tun_file connects an open character device to a tuntap netdevice. It |
92d4ea6e | 148 | * also contains all socket related structures (except sock_fprog and tap_filter) |
54f968d6 JW |
149 | * to serve as one transmit queue for tuntap device. The sock_fprog and |
150 | * tap_filter were kept in tun_struct since they were used for filtering for the | |
36fe8c09 | 151 | * netdevice not for a specific queue (at least I didn't see the requirement for |
54f968d6 | 152 | * this). |
6e914fc7 JW |
153 | * |
154 | * RCU usage: | |
36fe8c09 | 155 | * The tun_file and tun_struct are loosely coupled, the pointer from one to the |
6e914fc7 | 156 | * other can only be read while rcu_read_lock or rtnl_lock is held. |
54f968d6 | 157 | */ |
631ab46b | 158 | struct tun_file { |
54f968d6 JW |
159 | struct sock sk; |
160 | struct socket socket; | |
161 | struct socket_wq wq; | |
6e914fc7 | 162 | struct tun_struct __rcu *tun; |
54f968d6 JW |
163 | struct fasync_struct *fasync; |
164 | /* only used for fasnyc */ | |
165 | unsigned int flags; | |
fb7589a1 PE |
166 | union { |
167 | u16 queue_index; | |
168 | unsigned int ifindex; | |
169 | }; | |
4008e97f JW |
170 | struct list_head next; |
171 | struct tun_struct *detached; | |
1576d986 | 172 | struct skb_array tx_array; |
631ab46b EB |
173 | }; |
174 | ||
96442e42 JW |
175 | struct tun_flow_entry { |
176 | struct hlist_node hash_link; | |
177 | struct rcu_head rcu; | |
178 | struct tun_struct *tun; | |
179 | ||
180 | u32 rxhash; | |
9bc88939 | 181 | u32 rps_rxhash; |
96442e42 JW |
182 | int queue_index; |
183 | unsigned long updated; | |
184 | }; | |
185 | ||
186 | #define TUN_NUM_FLOW_ENTRIES 1024 | |
187 | ||
54f968d6 | 188 | /* Since the socket were moved to tun_file, to preserve the behavior of persist |
36fe8c09 | 189 | * device, socket filter, sndbuf and vnet header size were restore when the |
54f968d6 JW |
190 | * file were attached to a persist device. |
191 | */ | |
14daa021 | 192 | struct tun_struct { |
c8d68e6b JW |
193 | struct tun_file __rcu *tfiles[MAX_TAP_QUEUES]; |
194 | unsigned int numqueues; | |
f271b2cc | 195 | unsigned int flags; |
0625c883 EB |
196 | kuid_t owner; |
197 | kgid_t group; | |
14daa021 | 198 | |
14daa021 | 199 | struct net_device *dev; |
c8f44aff | 200 | netdev_features_t set_features; |
88255375 | 201 | #define TUN_USER_FEATURES (NETIF_F_HW_CSUM|NETIF_F_TSO_ECN|NETIF_F_TSO| \ |
e3e3c423 | 202 | NETIF_F_TSO6|NETIF_F_UFO) |
d9d52b51 | 203 | |
eaea34b2 | 204 | int align; |
d9d52b51 | 205 | int vnet_hdr_sz; |
54f968d6 JW |
206 | int sndbuf; |
207 | struct tap_filter txflt; | |
208 | struct sock_fprog fprog; | |
209 | /* protected by rtnl lock */ | |
210 | bool filter_attached; | |
14daa021 RR |
211 | #ifdef TUN_DEBUG |
212 | int debug; | |
1da177e4 | 213 | #endif |
96442e42 | 214 | spinlock_t lock; |
96442e42 JW |
215 | struct hlist_head flows[TUN_NUM_FLOW_ENTRIES]; |
216 | struct timer_list flow_gc_timer; | |
217 | unsigned long ageing_time; | |
4008e97f JW |
218 | unsigned int numdisabled; |
219 | struct list_head disabled; | |
5dbbaf2d | 220 | void *security; |
b8732fb7 | 221 | u32 flow_count; |
5503fcec | 222 | u32 rx_batched; |
608b9977 | 223 | struct tun_pcpu_stats __percpu *pcpu_stats; |
14daa021 | 224 | }; |
1da177e4 | 225 | |
8b8e658b GK |
226 | #ifdef CONFIG_TUN_VNET_CROSS_LE |
227 | static inline bool tun_legacy_is_little_endian(struct tun_struct *tun) | |
228 | { | |
229 | return tun->flags & TUN_VNET_BE ? false : | |
230 | virtio_legacy_is_little_endian(); | |
231 | } | |
232 | ||
233 | static long tun_get_vnet_be(struct tun_struct *tun, int __user *argp) | |
234 | { | |
235 | int be = !!(tun->flags & TUN_VNET_BE); | |
236 | ||
237 | if (put_user(be, argp)) | |
238 | return -EFAULT; | |
239 | ||
240 | return 0; | |
241 | } | |
242 | ||
243 | static long tun_set_vnet_be(struct tun_struct *tun, int __user *argp) | |
244 | { | |
245 | int be; | |
246 | ||
247 | if (get_user(be, argp)) | |
248 | return -EFAULT; | |
249 | ||
250 | if (be) | |
251 | tun->flags |= TUN_VNET_BE; | |
252 | else | |
253 | tun->flags &= ~TUN_VNET_BE; | |
254 | ||
255 | return 0; | |
256 | } | |
257 | #else | |
258 | static inline bool tun_legacy_is_little_endian(struct tun_struct *tun) | |
259 | { | |
260 | return virtio_legacy_is_little_endian(); | |
261 | } | |
262 | ||
263 | static long tun_get_vnet_be(struct tun_struct *tun, int __user *argp) | |
264 | { | |
265 | return -EINVAL; | |
266 | } | |
267 | ||
268 | static long tun_set_vnet_be(struct tun_struct *tun, int __user *argp) | |
269 | { | |
270 | return -EINVAL; | |
271 | } | |
272 | #endif /* CONFIG_TUN_VNET_CROSS_LE */ | |
273 | ||
25bd55bb GK |
274 | static inline bool tun_is_little_endian(struct tun_struct *tun) |
275 | { | |
7d824109 | 276 | return tun->flags & TUN_VNET_LE || |
8b8e658b | 277 | tun_legacy_is_little_endian(tun); |
25bd55bb GK |
278 | } |
279 | ||
56f0dcc5 MT |
280 | static inline u16 tun16_to_cpu(struct tun_struct *tun, __virtio16 val) |
281 | { | |
25bd55bb | 282 | return __virtio16_to_cpu(tun_is_little_endian(tun), val); |
56f0dcc5 MT |
283 | } |
284 | ||
285 | static inline __virtio16 cpu_to_tun16(struct tun_struct *tun, u16 val) | |
286 | { | |
25bd55bb | 287 | return __cpu_to_virtio16(tun_is_little_endian(tun), val); |
56f0dcc5 MT |
288 | } |
289 | ||
96442e42 JW |
290 | static inline u32 tun_hashfn(u32 rxhash) |
291 | { | |
292 | return rxhash & 0x3ff; | |
293 | } | |
294 | ||
295 | static struct tun_flow_entry *tun_flow_find(struct hlist_head *head, u32 rxhash) | |
296 | { | |
297 | struct tun_flow_entry *e; | |
96442e42 | 298 | |
b67bfe0d | 299 | hlist_for_each_entry_rcu(e, head, hash_link) { |
96442e42 JW |
300 | if (e->rxhash == rxhash) |
301 | return e; | |
302 | } | |
303 | return NULL; | |
304 | } | |
305 | ||
306 | static struct tun_flow_entry *tun_flow_create(struct tun_struct *tun, | |
307 | struct hlist_head *head, | |
308 | u32 rxhash, u16 queue_index) | |
309 | { | |
9fdc6bef ED |
310 | struct tun_flow_entry *e = kmalloc(sizeof(*e), GFP_ATOMIC); |
311 | ||
96442e42 JW |
312 | if (e) { |
313 | tun_debug(KERN_INFO, tun, "create flow: hash %u index %u\n", | |
314 | rxhash, queue_index); | |
315 | e->updated = jiffies; | |
316 | e->rxhash = rxhash; | |
9bc88939 | 317 | e->rps_rxhash = 0; |
96442e42 JW |
318 | e->queue_index = queue_index; |
319 | e->tun = tun; | |
320 | hlist_add_head_rcu(&e->hash_link, head); | |
b8732fb7 | 321 | ++tun->flow_count; |
96442e42 JW |
322 | } |
323 | return e; | |
324 | } | |
325 | ||
96442e42 JW |
326 | static void tun_flow_delete(struct tun_struct *tun, struct tun_flow_entry *e) |
327 | { | |
328 | tun_debug(KERN_INFO, tun, "delete flow: hash %u index %u\n", | |
329 | e->rxhash, e->queue_index); | |
330 | hlist_del_rcu(&e->hash_link); | |
9fdc6bef | 331 | kfree_rcu(e, rcu); |
b8732fb7 | 332 | --tun->flow_count; |
96442e42 JW |
333 | } |
334 | ||
335 | static void tun_flow_flush(struct tun_struct *tun) | |
336 | { | |
337 | int i; | |
338 | ||
339 | spin_lock_bh(&tun->lock); | |
340 | for (i = 0; i < TUN_NUM_FLOW_ENTRIES; i++) { | |
341 | struct tun_flow_entry *e; | |
b67bfe0d | 342 | struct hlist_node *n; |
96442e42 | 343 | |
b67bfe0d | 344 | hlist_for_each_entry_safe(e, n, &tun->flows[i], hash_link) |
96442e42 JW |
345 | tun_flow_delete(tun, e); |
346 | } | |
347 | spin_unlock_bh(&tun->lock); | |
348 | } | |
349 | ||
350 | static void tun_flow_delete_by_queue(struct tun_struct *tun, u16 queue_index) | |
351 | { | |
352 | int i; | |
353 | ||
354 | spin_lock_bh(&tun->lock); | |
355 | for (i = 0; i < TUN_NUM_FLOW_ENTRIES; i++) { | |
356 | struct tun_flow_entry *e; | |
b67bfe0d | 357 | struct hlist_node *n; |
96442e42 | 358 | |
b67bfe0d | 359 | hlist_for_each_entry_safe(e, n, &tun->flows[i], hash_link) { |
96442e42 JW |
360 | if (e->queue_index == queue_index) |
361 | tun_flow_delete(tun, e); | |
362 | } | |
363 | } | |
364 | spin_unlock_bh(&tun->lock); | |
365 | } | |
366 | ||
367 | static void tun_flow_cleanup(unsigned long data) | |
368 | { | |
369 | struct tun_struct *tun = (struct tun_struct *)data; | |
370 | unsigned long delay = tun->ageing_time; | |
371 | unsigned long next_timer = jiffies + delay; | |
372 | unsigned long count = 0; | |
373 | int i; | |
374 | ||
375 | tun_debug(KERN_INFO, tun, "tun_flow_cleanup\n"); | |
376 | ||
377 | spin_lock_bh(&tun->lock); | |
378 | for (i = 0; i < TUN_NUM_FLOW_ENTRIES; i++) { | |
379 | struct tun_flow_entry *e; | |
b67bfe0d | 380 | struct hlist_node *n; |
96442e42 | 381 | |
b67bfe0d | 382 | hlist_for_each_entry_safe(e, n, &tun->flows[i], hash_link) { |
96442e42 JW |
383 | unsigned long this_timer; |
384 | count++; | |
385 | this_timer = e->updated + delay; | |
386 | if (time_before_eq(this_timer, jiffies)) | |
387 | tun_flow_delete(tun, e); | |
388 | else if (time_before(this_timer, next_timer)) | |
389 | next_timer = this_timer; | |
390 | } | |
391 | } | |
392 | ||
393 | if (count) | |
394 | mod_timer(&tun->flow_gc_timer, round_jiffies_up(next_timer)); | |
395 | spin_unlock_bh(&tun->lock); | |
396 | } | |
397 | ||
49974420 | 398 | static void tun_flow_update(struct tun_struct *tun, u32 rxhash, |
9e85722d | 399 | struct tun_file *tfile) |
96442e42 JW |
400 | { |
401 | struct hlist_head *head; | |
402 | struct tun_flow_entry *e; | |
403 | unsigned long delay = tun->ageing_time; | |
9e85722d | 404 | u16 queue_index = tfile->queue_index; |
96442e42 JW |
405 | |
406 | if (!rxhash) | |
407 | return; | |
408 | else | |
409 | head = &tun->flows[tun_hashfn(rxhash)]; | |
410 | ||
411 | rcu_read_lock(); | |
412 | ||
9e85722d JW |
413 | /* We may get a very small possibility of OOO during switching, not |
414 | * worth to optimize.*/ | |
415 | if (tun->numqueues == 1 || tfile->detached) | |
96442e42 JW |
416 | goto unlock; |
417 | ||
418 | e = tun_flow_find(head, rxhash); | |
419 | if (likely(e)) { | |
420 | /* TODO: keep queueing to old queue until it's empty? */ | |
421 | e->queue_index = queue_index; | |
422 | e->updated = jiffies; | |
9bc88939 | 423 | sock_rps_record_flow_hash(e->rps_rxhash); |
96442e42 JW |
424 | } else { |
425 | spin_lock_bh(&tun->lock); | |
b8732fb7 JW |
426 | if (!tun_flow_find(head, rxhash) && |
427 | tun->flow_count < MAX_TAP_FLOWS) | |
96442e42 JW |
428 | tun_flow_create(tun, head, rxhash, queue_index); |
429 | ||
430 | if (!timer_pending(&tun->flow_gc_timer)) | |
431 | mod_timer(&tun->flow_gc_timer, | |
432 | round_jiffies_up(jiffies + delay)); | |
433 | spin_unlock_bh(&tun->lock); | |
434 | } | |
435 | ||
436 | unlock: | |
437 | rcu_read_unlock(); | |
438 | } | |
439 | ||
9bc88939 TH |
440 | /** |
441 | * Save the hash received in the stack receive path and update the | |
442 | * flow_hash table accordingly. | |
443 | */ | |
444 | static inline void tun_flow_save_rps_rxhash(struct tun_flow_entry *e, u32 hash) | |
445 | { | |
567e4b79 | 446 | if (unlikely(e->rps_rxhash != hash)) |
9bc88939 | 447 | e->rps_rxhash = hash; |
9bc88939 TH |
448 | } |
449 | ||
c8d68e6b | 450 | /* We try to identify a flow through its rxhash first. The reason that |
92d4ea6e | 451 | * we do not check rxq no. is because some cards(e.g 82599), chooses |
c8d68e6b JW |
452 | * the rxq based on the txq where the last packet of the flow comes. As |
453 | * the userspace application move between processors, we may get a | |
454 | * different rxq no. here. If we could not get rxhash, then we would | |
455 | * hope the rxq no. may help here. | |
456 | */ | |
f663dd9a | 457 | static u16 tun_select_queue(struct net_device *dev, struct sk_buff *skb, |
99932d4f | 458 | void *accel_priv, select_queue_fallback_t fallback) |
c8d68e6b JW |
459 | { |
460 | struct tun_struct *tun = netdev_priv(dev); | |
96442e42 | 461 | struct tun_flow_entry *e; |
c8d68e6b JW |
462 | u32 txq = 0; |
463 | u32 numqueues = 0; | |
464 | ||
465 | rcu_read_lock(); | |
92bb73ea | 466 | numqueues = ACCESS_ONCE(tun->numqueues); |
c8d68e6b | 467 | |
3958afa1 | 468 | txq = skb_get_hash(skb); |
c8d68e6b | 469 | if (txq) { |
96442e42 | 470 | e = tun_flow_find(&tun->flows[tun_hashfn(txq)], txq); |
9bc88939 | 471 | if (e) { |
9bc88939 | 472 | tun_flow_save_rps_rxhash(e, txq); |
fbe4d456 | 473 | txq = e->queue_index; |
9bc88939 | 474 | } else |
96442e42 JW |
475 | /* use multiply and shift instead of expensive divide */ |
476 | txq = ((u64)txq * numqueues) >> 32; | |
c8d68e6b JW |
477 | } else if (likely(skb_rx_queue_recorded(skb))) { |
478 | txq = skb_get_rx_queue(skb); | |
479 | while (unlikely(txq >= numqueues)) | |
480 | txq -= numqueues; | |
481 | } | |
482 | ||
483 | rcu_read_unlock(); | |
484 | return txq; | |
485 | } | |
486 | ||
cde8b15f JW |
487 | static inline bool tun_not_capable(struct tun_struct *tun) |
488 | { | |
489 | const struct cred *cred = current_cred(); | |
c260b772 | 490 | struct net *net = dev_net(tun->dev); |
cde8b15f JW |
491 | |
492 | return ((uid_valid(tun->owner) && !uid_eq(cred->euid, tun->owner)) || | |
493 | (gid_valid(tun->group) && !in_egroup_p(tun->group))) && | |
c260b772 | 494 | !ns_capable(net->user_ns, CAP_NET_ADMIN); |
cde8b15f JW |
495 | } |
496 | ||
c8d68e6b JW |
497 | static void tun_set_real_num_queues(struct tun_struct *tun) |
498 | { | |
499 | netif_set_real_num_tx_queues(tun->dev, tun->numqueues); | |
500 | netif_set_real_num_rx_queues(tun->dev, tun->numqueues); | |
501 | } | |
502 | ||
4008e97f JW |
503 | static void tun_disable_queue(struct tun_struct *tun, struct tun_file *tfile) |
504 | { | |
505 | tfile->detached = tun; | |
506 | list_add_tail(&tfile->next, &tun->disabled); | |
507 | ++tun->numdisabled; | |
508 | } | |
509 | ||
d32649d1 | 510 | static struct tun_struct *tun_enable_queue(struct tun_file *tfile) |
4008e97f JW |
511 | { |
512 | struct tun_struct *tun = tfile->detached; | |
513 | ||
514 | tfile->detached = NULL; | |
515 | list_del_init(&tfile->next); | |
516 | --tun->numdisabled; | |
517 | return tun; | |
518 | } | |
519 | ||
4bfb0513 JW |
520 | static void tun_queue_purge(struct tun_file *tfile) |
521 | { | |
1576d986 JW |
522 | struct sk_buff *skb; |
523 | ||
524 | while ((skb = skb_array_consume(&tfile->tx_array)) != NULL) | |
525 | kfree_skb(skb); | |
526 | ||
5503fcec | 527 | skb_queue_purge(&tfile->sk.sk_write_queue); |
4bfb0513 JW |
528 | skb_queue_purge(&tfile->sk.sk_error_queue); |
529 | } | |
530 | ||
c8d68e6b JW |
531 | static void __tun_detach(struct tun_file *tfile, bool clean) |
532 | { | |
533 | struct tun_file *ntfile; | |
534 | struct tun_struct *tun; | |
c8d68e6b | 535 | |
b8deabd3 JW |
536 | tun = rtnl_dereference(tfile->tun); |
537 | ||
9e85722d | 538 | if (tun && !tfile->detached) { |
c8d68e6b JW |
539 | u16 index = tfile->queue_index; |
540 | BUG_ON(index >= tun->numqueues); | |
c8d68e6b JW |
541 | |
542 | rcu_assign_pointer(tun->tfiles[index], | |
543 | tun->tfiles[tun->numqueues - 1]); | |
b8deabd3 | 544 | ntfile = rtnl_dereference(tun->tfiles[index]); |
c8d68e6b JW |
545 | ntfile->queue_index = index; |
546 | ||
547 | --tun->numqueues; | |
9e85722d | 548 | if (clean) { |
c956674b | 549 | RCU_INIT_POINTER(tfile->tun, NULL); |
4008e97f | 550 | sock_put(&tfile->sk); |
9e85722d | 551 | } else |
4008e97f | 552 | tun_disable_queue(tun, tfile); |
c8d68e6b JW |
553 | |
554 | synchronize_net(); | |
96442e42 | 555 | tun_flow_delete_by_queue(tun, tun->numqueues + 1); |
c8d68e6b | 556 | /* Drop read queue */ |
4bfb0513 | 557 | tun_queue_purge(tfile); |
c8d68e6b | 558 | tun_set_real_num_queues(tun); |
dd38bd85 | 559 | } else if (tfile->detached && clean) { |
4008e97f | 560 | tun = tun_enable_queue(tfile); |
dd38bd85 JW |
561 | sock_put(&tfile->sk); |
562 | } | |
c8d68e6b JW |
563 | |
564 | if (clean) { | |
af668b3c MT |
565 | if (tun && tun->numqueues == 0 && tun->numdisabled == 0) { |
566 | netif_carrier_off(tun->dev); | |
567 | ||
40630b82 | 568 | if (!(tun->flags & IFF_PERSIST) && |
af668b3c | 569 | tun->dev->reg_state == NETREG_REGISTERED) |
4008e97f | 570 | unregister_netdevice(tun->dev); |
af668b3c | 571 | } |
1576d986 JW |
572 | if (tun) |
573 | skb_array_cleanup(&tfile->tx_array); | |
140e807d | 574 | sock_put(&tfile->sk); |
c8d68e6b JW |
575 | } |
576 | } | |
577 | ||
578 | static void tun_detach(struct tun_file *tfile, bool clean) | |
579 | { | |
580 | rtnl_lock(); | |
581 | __tun_detach(tfile, clean); | |
582 | rtnl_unlock(); | |
583 | } | |
584 | ||
585 | static void tun_detach_all(struct net_device *dev) | |
586 | { | |
587 | struct tun_struct *tun = netdev_priv(dev); | |
4008e97f | 588 | struct tun_file *tfile, *tmp; |
c8d68e6b JW |
589 | int i, n = tun->numqueues; |
590 | ||
591 | for (i = 0; i < n; i++) { | |
b8deabd3 | 592 | tfile = rtnl_dereference(tun->tfiles[i]); |
c8d68e6b | 593 | BUG_ON(!tfile); |
addf8fc4 | 594 | tfile->socket.sk->sk_shutdown = RCV_SHUTDOWN; |
9e641bdc | 595 | tfile->socket.sk->sk_data_ready(tfile->socket.sk); |
c956674b | 596 | RCU_INIT_POINTER(tfile->tun, NULL); |
c8d68e6b JW |
597 | --tun->numqueues; |
598 | } | |
9e85722d | 599 | list_for_each_entry(tfile, &tun->disabled, next) { |
addf8fc4 | 600 | tfile->socket.sk->sk_shutdown = RCV_SHUTDOWN; |
9e641bdc | 601 | tfile->socket.sk->sk_data_ready(tfile->socket.sk); |
c956674b | 602 | RCU_INIT_POINTER(tfile->tun, NULL); |
9e85722d | 603 | } |
c8d68e6b JW |
604 | BUG_ON(tun->numqueues != 0); |
605 | ||
606 | synchronize_net(); | |
607 | for (i = 0; i < n; i++) { | |
b8deabd3 | 608 | tfile = rtnl_dereference(tun->tfiles[i]); |
c8d68e6b | 609 | /* Drop read queue */ |
4bfb0513 | 610 | tun_queue_purge(tfile); |
c8d68e6b JW |
611 | sock_put(&tfile->sk); |
612 | } | |
4008e97f JW |
613 | list_for_each_entry_safe(tfile, tmp, &tun->disabled, next) { |
614 | tun_enable_queue(tfile); | |
4bfb0513 | 615 | tun_queue_purge(tfile); |
4008e97f JW |
616 | sock_put(&tfile->sk); |
617 | } | |
618 | BUG_ON(tun->numdisabled != 0); | |
dd38bd85 | 619 | |
40630b82 | 620 | if (tun->flags & IFF_PERSIST) |
dd38bd85 | 621 | module_put(THIS_MODULE); |
c8d68e6b JW |
622 | } |
623 | ||
849c9b6f | 624 | static int tun_attach(struct tun_struct *tun, struct file *file, bool skip_filter) |
a7385ba2 | 625 | { |
631ab46b | 626 | struct tun_file *tfile = file->private_data; |
1576d986 | 627 | struct net_device *dev = tun->dev; |
38231b7a | 628 | int err; |
a7385ba2 | 629 | |
5dbbaf2d PM |
630 | err = security_tun_dev_attach(tfile->socket.sk, tun->security); |
631 | if (err < 0) | |
632 | goto out; | |
633 | ||
38231b7a | 634 | err = -EINVAL; |
9e85722d | 635 | if (rtnl_dereference(tfile->tun) && !tfile->detached) |
38231b7a EB |
636 | goto out; |
637 | ||
638 | err = -EBUSY; | |
40630b82 | 639 | if (!(tun->flags & IFF_MULTI_QUEUE) && tun->numqueues == 1) |
c8d68e6b JW |
640 | goto out; |
641 | ||
642 | err = -E2BIG; | |
4008e97f JW |
643 | if (!tfile->detached && |
644 | tun->numqueues + tun->numdisabled == MAX_TAP_QUEUES) | |
38231b7a EB |
645 | goto out; |
646 | ||
647 | err = 0; | |
54f968d6 | 648 | |
92d4ea6e | 649 | /* Re-attach the filter to persist device */ |
849c9b6f | 650 | if (!skip_filter && (tun->filter_attached == true)) { |
8ced425e HFS |
651 | lock_sock(tfile->socket.sk); |
652 | err = sk_attach_filter(&tun->fprog, tfile->socket.sk); | |
653 | release_sock(tfile->socket.sk); | |
54f968d6 JW |
654 | if (!err) |
655 | goto out; | |
656 | } | |
1576d986 JW |
657 | |
658 | if (!tfile->detached && | |
659 | skb_array_init(&tfile->tx_array, dev->tx_queue_len, GFP_KERNEL)) { | |
660 | err = -ENOMEM; | |
661 | goto out; | |
662 | } | |
663 | ||
c8d68e6b | 664 | tfile->queue_index = tun->numqueues; |
addf8fc4 | 665 | tfile->socket.sk->sk_shutdown &= ~RCV_SHUTDOWN; |
6e914fc7 | 666 | rcu_assign_pointer(tfile->tun, tun); |
c8d68e6b | 667 | rcu_assign_pointer(tun->tfiles[tun->numqueues], tfile); |
c8d68e6b | 668 | tun->numqueues++; |
a7385ba2 | 669 | |
4008e97f JW |
670 | if (tfile->detached) |
671 | tun_enable_queue(tfile); | |
672 | else | |
673 | sock_hold(&tfile->sk); | |
674 | ||
c8d68e6b | 675 | tun_set_real_num_queues(tun); |
a7385ba2 | 676 | |
c8d68e6b JW |
677 | /* device is allowed to go away first, so no need to hold extra |
678 | * refcnt. | |
679 | */ | |
680 | ||
681 | out: | |
682 | return err; | |
631ab46b EB |
683 | } |
684 | ||
685 | static struct tun_struct *__tun_get(struct tun_file *tfile) | |
686 | { | |
6e914fc7 | 687 | struct tun_struct *tun; |
c70f1829 | 688 | |
6e914fc7 JW |
689 | rcu_read_lock(); |
690 | tun = rcu_dereference(tfile->tun); | |
691 | if (tun) | |
692 | dev_hold(tun->dev); | |
693 | rcu_read_unlock(); | |
c70f1829 EB |
694 | |
695 | return tun; | |
631ab46b EB |
696 | } |
697 | ||
698 | static struct tun_struct *tun_get(struct file *file) | |
699 | { | |
700 | return __tun_get(file->private_data); | |
701 | } | |
702 | ||
703 | static void tun_put(struct tun_struct *tun) | |
704 | { | |
6e914fc7 | 705 | dev_put(tun->dev); |
631ab46b EB |
706 | } |
707 | ||
6b8a66ee | 708 | /* TAP filtering */ |
f271b2cc MK |
709 | static void addr_hash_set(u32 *mask, const u8 *addr) |
710 | { | |
711 | int n = ether_crc(ETH_ALEN, addr) >> 26; | |
712 | mask[n >> 5] |= (1 << (n & 31)); | |
713 | } | |
714 | ||
715 | static unsigned int addr_hash_test(const u32 *mask, const u8 *addr) | |
716 | { | |
717 | int n = ether_crc(ETH_ALEN, addr) >> 26; | |
718 | return mask[n >> 5] & (1 << (n & 31)); | |
719 | } | |
720 | ||
721 | static int update_filter(struct tap_filter *filter, void __user *arg) | |
722 | { | |
723 | struct { u8 u[ETH_ALEN]; } *addr; | |
724 | struct tun_filter uf; | |
725 | int err, alen, n, nexact; | |
726 | ||
727 | if (copy_from_user(&uf, arg, sizeof(uf))) | |
728 | return -EFAULT; | |
729 | ||
730 | if (!uf.count) { | |
731 | /* Disabled */ | |
732 | filter->count = 0; | |
733 | return 0; | |
734 | } | |
735 | ||
736 | alen = ETH_ALEN * uf.count; | |
28e8190d ME |
737 | addr = memdup_user(arg + sizeof(uf), alen); |
738 | if (IS_ERR(addr)) | |
739 | return PTR_ERR(addr); | |
f271b2cc MK |
740 | |
741 | /* The filter is updated without holding any locks. Which is | |
742 | * perfectly safe. We disable it first and in the worst | |
743 | * case we'll accept a few undesired packets. */ | |
744 | filter->count = 0; | |
745 | wmb(); | |
746 | ||
747 | /* Use first set of addresses as an exact filter */ | |
748 | for (n = 0; n < uf.count && n < FLT_EXACT_COUNT; n++) | |
749 | memcpy(filter->addr[n], addr[n].u, ETH_ALEN); | |
750 | ||
751 | nexact = n; | |
752 | ||
cfbf84fc AW |
753 | /* Remaining multicast addresses are hashed, |
754 | * unicast will leave the filter disabled. */ | |
f271b2cc | 755 | memset(filter->mask, 0, sizeof(filter->mask)); |
cfbf84fc AW |
756 | for (; n < uf.count; n++) { |
757 | if (!is_multicast_ether_addr(addr[n].u)) { | |
758 | err = 0; /* no filter */ | |
3b8d2a69 | 759 | goto free_addr; |
cfbf84fc | 760 | } |
f271b2cc | 761 | addr_hash_set(filter->mask, addr[n].u); |
cfbf84fc | 762 | } |
f271b2cc MK |
763 | |
764 | /* For ALLMULTI just set the mask to all ones. | |
765 | * This overrides the mask populated above. */ | |
766 | if ((uf.flags & TUN_FLT_ALLMULTI)) | |
767 | memset(filter->mask, ~0, sizeof(filter->mask)); | |
768 | ||
769 | /* Now enable the filter */ | |
770 | wmb(); | |
771 | filter->count = nexact; | |
772 | ||
773 | /* Return the number of exact filters */ | |
774 | err = nexact; | |
3b8d2a69 | 775 | free_addr: |
f271b2cc MK |
776 | kfree(addr); |
777 | return err; | |
778 | } | |
779 | ||
780 | /* Returns: 0 - drop, !=0 - accept */ | |
781 | static int run_filter(struct tap_filter *filter, const struct sk_buff *skb) | |
782 | { | |
783 | /* Cannot use eth_hdr(skb) here because skb_mac_hdr() is incorrect | |
784 | * at this point. */ | |
785 | struct ethhdr *eh = (struct ethhdr *) skb->data; | |
786 | int i; | |
787 | ||
788 | /* Exact match */ | |
789 | for (i = 0; i < filter->count; i++) | |
2e42e474 | 790 | if (ether_addr_equal(eh->h_dest, filter->addr[i])) |
f271b2cc MK |
791 | return 1; |
792 | ||
793 | /* Inexact match (multicast only) */ | |
794 | if (is_multicast_ether_addr(eh->h_dest)) | |
795 | return addr_hash_test(filter->mask, eh->h_dest); | |
796 | ||
797 | return 0; | |
798 | } | |
799 | ||
800 | /* | |
801 | * Checks whether the packet is accepted or not. | |
802 | * Returns: 0 - drop, !=0 - accept | |
803 | */ | |
804 | static int check_filter(struct tap_filter *filter, const struct sk_buff *skb) | |
805 | { | |
806 | if (!filter->count) | |
807 | return 1; | |
808 | ||
809 | return run_filter(filter, skb); | |
810 | } | |
811 | ||
1da177e4 LT |
812 | /* Network device part of the driver */ |
813 | ||
7282d491 | 814 | static const struct ethtool_ops tun_ethtool_ops; |
1da177e4 | 815 | |
c70f1829 EB |
816 | /* Net device detach from fd. */ |
817 | static void tun_net_uninit(struct net_device *dev) | |
818 | { | |
c8d68e6b | 819 | tun_detach_all(dev); |
c70f1829 EB |
820 | } |
821 | ||
1da177e4 LT |
822 | /* Net device open. */ |
823 | static int tun_net_open(struct net_device *dev) | |
824 | { | |
b20e2d54 HFS |
825 | struct tun_struct *tun = netdev_priv(dev); |
826 | int i; | |
827 | ||
c8d68e6b | 828 | netif_tx_start_all_queues(dev); |
b20e2d54 HFS |
829 | |
830 | for (i = 0; i < tun->numqueues; i++) { | |
831 | struct tun_file *tfile; | |
832 | ||
833 | tfile = rtnl_dereference(tun->tfiles[i]); | |
834 | tfile->socket.sk->sk_write_space(tfile->socket.sk); | |
835 | } | |
836 | ||
1da177e4 LT |
837 | return 0; |
838 | } | |
839 | ||
840 | /* Net device close. */ | |
841 | static int tun_net_close(struct net_device *dev) | |
842 | { | |
c8d68e6b | 843 | netif_tx_stop_all_queues(dev); |
1da177e4 LT |
844 | return 0; |
845 | } | |
846 | ||
847 | /* Net device start xmit */ | |
424efe9c | 848 | static netdev_tx_t tun_net_xmit(struct sk_buff *skb, struct net_device *dev) |
1da177e4 LT |
849 | { |
850 | struct tun_struct *tun = netdev_priv(dev); | |
c8d68e6b | 851 | int txq = skb->queue_mapping; |
6e914fc7 | 852 | struct tun_file *tfile; |
fa35864e | 853 | u32 numqueues = 0; |
1da177e4 | 854 | |
6e914fc7 | 855 | rcu_read_lock(); |
c8d68e6b | 856 | tfile = rcu_dereference(tun->tfiles[txq]); |
fa35864e | 857 | numqueues = ACCESS_ONCE(tun->numqueues); |
c8d68e6b | 858 | |
1da177e4 | 859 | /* Drop packet if interface is not attached */ |
fa35864e | 860 | if (txq >= numqueues) |
1da177e4 LT |
861 | goto drop; |
862 | ||
3df97ba8 JW |
863 | #ifdef CONFIG_RPS |
864 | if (numqueues == 1 && static_key_false(&rps_needed)) { | |
9bc88939 TH |
865 | /* Select queue was not called for the skbuff, so we extract the |
866 | * RPS hash and save it into the flow_table here. | |
867 | */ | |
868 | __u32 rxhash; | |
869 | ||
870 | rxhash = skb_get_hash(skb); | |
871 | if (rxhash) { | |
872 | struct tun_flow_entry *e; | |
873 | e = tun_flow_find(&tun->flows[tun_hashfn(rxhash)], | |
874 | rxhash); | |
875 | if (e) | |
876 | tun_flow_save_rps_rxhash(e, rxhash); | |
877 | } | |
878 | } | |
3df97ba8 | 879 | #endif |
9bc88939 | 880 | |
6e914fc7 JW |
881 | tun_debug(KERN_INFO, tun, "tun_net_xmit %d\n", skb->len); |
882 | ||
c8d68e6b JW |
883 | BUG_ON(!tfile); |
884 | ||
f271b2cc MK |
885 | /* Drop if the filter does not like it. |
886 | * This is a noop if the filter is disabled. | |
887 | * Filter can be enabled only for the TAP devices. */ | |
888 | if (!check_filter(&tun->txflt, skb)) | |
889 | goto drop; | |
890 | ||
54f968d6 JW |
891 | if (tfile->socket.sk->sk_filter && |
892 | sk_filter(tfile->socket.sk, skb)) | |
99405162 MT |
893 | goto drop; |
894 | ||
7bf66305 JW |
895 | if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC))) |
896 | goto drop; | |
897 | ||
7b996243 | 898 | skb_tx_timestamp(skb); |
eda29772 | 899 | |
0110d6f2 | 900 | /* Orphan the skb - required as we might hang on to it |
7bf66305 JW |
901 | * for indefinite time. |
902 | */ | |
0110d6f2 MT |
903 | skb_orphan(skb); |
904 | ||
f8af75f3 ED |
905 | nf_reset(skb); |
906 | ||
1576d986 JW |
907 | if (skb_array_produce(&tfile->tx_array, skb)) |
908 | goto drop; | |
1da177e4 LT |
909 | |
910 | /* Notify and wake up reader process */ | |
54f968d6 JW |
911 | if (tfile->flags & TUN_FASYNC) |
912 | kill_fasync(&tfile->fasync, SIGIO, POLL_IN); | |
9e641bdc | 913 | tfile->socket.sk->sk_data_ready(tfile->socket.sk); |
6e914fc7 JW |
914 | |
915 | rcu_read_unlock(); | |
6ed10654 | 916 | return NETDEV_TX_OK; |
1da177e4 LT |
917 | |
918 | drop: | |
608b9977 | 919 | this_cpu_inc(tun->pcpu_stats->tx_dropped); |
149d36f7 | 920 | skb_tx_error(skb); |
1da177e4 | 921 | kfree_skb(skb); |
6e914fc7 | 922 | rcu_read_unlock(); |
baeababb | 923 | return NET_XMIT_DROP; |
1da177e4 LT |
924 | } |
925 | ||
f271b2cc | 926 | static void tun_net_mclist(struct net_device *dev) |
1da177e4 | 927 | { |
f271b2cc MK |
928 | /* |
929 | * This callback is supposed to deal with mc filter in | |
930 | * _rx_ path and has nothing to do with the _tx_ path. | |
931 | * In rx path we always accept everything userspace gives us. | |
932 | */ | |
1da177e4 LT |
933 | } |
934 | ||
c8f44aff MM |
935 | static netdev_features_t tun_net_fix_features(struct net_device *dev, |
936 | netdev_features_t features) | |
88255375 MM |
937 | { |
938 | struct tun_struct *tun = netdev_priv(dev); | |
939 | ||
940 | return (features & tun->set_features) | (features & ~TUN_USER_FEATURES); | |
941 | } | |
bebd097a NH |
942 | #ifdef CONFIG_NET_POLL_CONTROLLER |
943 | static void tun_poll_controller(struct net_device *dev) | |
944 | { | |
945 | /* | |
946 | * Tun only receives frames when: | |
947 | * 1) the char device endpoint gets data from user space | |
948 | * 2) the tun socket gets a sendmsg call from user space | |
92d4ea6e | 949 | * Since both of those are synchronous operations, we are guaranteed |
bebd097a | 950 | * never to have pending data when we poll for it |
92d4ea6e | 951 | * so there is nothing to do here but return. |
bebd097a NH |
952 | * We need this though so netpoll recognizes us as an interface that |
953 | * supports polling, which enables bridge devices in virt setups to | |
954 | * still use netconsole | |
955 | */ | |
956 | return; | |
957 | } | |
958 | #endif | |
eaea34b2 PA |
959 | |
960 | static void tun_set_headroom(struct net_device *dev, int new_hr) | |
961 | { | |
962 | struct tun_struct *tun = netdev_priv(dev); | |
963 | ||
964 | if (new_hr < NET_SKB_PAD) | |
965 | new_hr = NET_SKB_PAD; | |
966 | ||
967 | tun->align = new_hr; | |
968 | } | |
969 | ||
bc1f4470 | 970 | static void |
608b9977 PA |
971 | tun_net_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats) |
972 | { | |
973 | u32 rx_dropped = 0, tx_dropped = 0, rx_frame_errors = 0; | |
974 | struct tun_struct *tun = netdev_priv(dev); | |
975 | struct tun_pcpu_stats *p; | |
976 | int i; | |
977 | ||
978 | for_each_possible_cpu(i) { | |
979 | u64 rxpackets, rxbytes, txpackets, txbytes; | |
980 | unsigned int start; | |
981 | ||
982 | p = per_cpu_ptr(tun->pcpu_stats, i); | |
983 | do { | |
984 | start = u64_stats_fetch_begin(&p->syncp); | |
985 | rxpackets = p->rx_packets; | |
986 | rxbytes = p->rx_bytes; | |
987 | txpackets = p->tx_packets; | |
988 | txbytes = p->tx_bytes; | |
989 | } while (u64_stats_fetch_retry(&p->syncp, start)); | |
990 | ||
991 | stats->rx_packets += rxpackets; | |
992 | stats->rx_bytes += rxbytes; | |
993 | stats->tx_packets += txpackets; | |
994 | stats->tx_bytes += txbytes; | |
995 | ||
996 | /* u32 counters */ | |
997 | rx_dropped += p->rx_dropped; | |
998 | rx_frame_errors += p->rx_frame_errors; | |
999 | tx_dropped += p->tx_dropped; | |
1000 | } | |
1001 | stats->rx_dropped = rx_dropped; | |
1002 | stats->rx_frame_errors = rx_frame_errors; | |
1003 | stats->tx_dropped = tx_dropped; | |
608b9977 PA |
1004 | } |
1005 | ||
758e43b7 | 1006 | static const struct net_device_ops tun_netdev_ops = { |
c70f1829 | 1007 | .ndo_uninit = tun_net_uninit, |
758e43b7 SH |
1008 | .ndo_open = tun_net_open, |
1009 | .ndo_stop = tun_net_close, | |
00829823 | 1010 | .ndo_start_xmit = tun_net_xmit, |
88255375 | 1011 | .ndo_fix_features = tun_net_fix_features, |
c8d68e6b | 1012 | .ndo_select_queue = tun_select_queue, |
bebd097a NH |
1013 | #ifdef CONFIG_NET_POLL_CONTROLLER |
1014 | .ndo_poll_controller = tun_poll_controller, | |
1015 | #endif | |
eaea34b2 | 1016 | .ndo_set_rx_headroom = tun_set_headroom, |
608b9977 | 1017 | .ndo_get_stats64 = tun_net_get_stats64, |
758e43b7 SH |
1018 | }; |
1019 | ||
1020 | static const struct net_device_ops tap_netdev_ops = { | |
c70f1829 | 1021 | .ndo_uninit = tun_net_uninit, |
758e43b7 SH |
1022 | .ndo_open = tun_net_open, |
1023 | .ndo_stop = tun_net_close, | |
00829823 | 1024 | .ndo_start_xmit = tun_net_xmit, |
88255375 | 1025 | .ndo_fix_features = tun_net_fix_features, |
afc4b13d | 1026 | .ndo_set_rx_mode = tun_net_mclist, |
758e43b7 SH |
1027 | .ndo_set_mac_address = eth_mac_addr, |
1028 | .ndo_validate_addr = eth_validate_addr, | |
c8d68e6b | 1029 | .ndo_select_queue = tun_select_queue, |
bebd097a NH |
1030 | #ifdef CONFIG_NET_POLL_CONTROLLER |
1031 | .ndo_poll_controller = tun_poll_controller, | |
1032 | #endif | |
5e52796a | 1033 | .ndo_features_check = passthru_features_check, |
eaea34b2 | 1034 | .ndo_set_rx_headroom = tun_set_headroom, |
608b9977 | 1035 | .ndo_get_stats64 = tun_net_get_stats64, |
758e43b7 SH |
1036 | }; |
1037 | ||
944a1376 | 1038 | static void tun_flow_init(struct tun_struct *tun) |
96442e42 JW |
1039 | { |
1040 | int i; | |
1041 | ||
96442e42 JW |
1042 | for (i = 0; i < TUN_NUM_FLOW_ENTRIES; i++) |
1043 | INIT_HLIST_HEAD(&tun->flows[i]); | |
1044 | ||
1045 | tun->ageing_time = TUN_FLOW_EXPIRE; | |
1046 | setup_timer(&tun->flow_gc_timer, tun_flow_cleanup, (unsigned long)tun); | |
1047 | mod_timer(&tun->flow_gc_timer, | |
1048 | round_jiffies_up(jiffies + tun->ageing_time)); | |
96442e42 JW |
1049 | } |
1050 | ||
1051 | static void tun_flow_uninit(struct tun_struct *tun) | |
1052 | { | |
1053 | del_timer_sync(&tun->flow_gc_timer); | |
1054 | tun_flow_flush(tun); | |
96442e42 JW |
1055 | } |
1056 | ||
91572088 JW |
1057 | #define MIN_MTU 68 |
1058 | #define MAX_MTU 65535 | |
1059 | ||
1da177e4 LT |
1060 | /* Initialize net device. */ |
1061 | static void tun_net_init(struct net_device *dev) | |
1062 | { | |
1063 | struct tun_struct *tun = netdev_priv(dev); | |
6aa20a22 | 1064 | |
1da177e4 | 1065 | switch (tun->flags & TUN_TYPE_MASK) { |
40630b82 | 1066 | case IFF_TUN: |
758e43b7 SH |
1067 | dev->netdev_ops = &tun_netdev_ops; |
1068 | ||
1da177e4 LT |
1069 | /* Point-to-Point TUN Device */ |
1070 | dev->hard_header_len = 0; | |
1071 | dev->addr_len = 0; | |
1072 | dev->mtu = 1500; | |
1073 | ||
1074 | /* Zero header length */ | |
6aa20a22 | 1075 | dev->type = ARPHRD_NONE; |
1da177e4 | 1076 | dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST; |
1da177e4 LT |
1077 | break; |
1078 | ||
40630b82 | 1079 | case IFF_TAP: |
7a0a9608 | 1080 | dev->netdev_ops = &tap_netdev_ops; |
1da177e4 | 1081 | /* Ethernet TAP Device */ |
1da177e4 | 1082 | ether_setup(dev); |
550fd08c | 1083 | dev->priv_flags &= ~IFF_TX_SKB_SHARING; |
a676847b | 1084 | dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; |
36226a8d | 1085 | |
f2cedb63 | 1086 | eth_hw_addr_random(dev); |
36226a8d | 1087 | |
1da177e4 LT |
1088 | break; |
1089 | } | |
91572088 JW |
1090 | |
1091 | dev->min_mtu = MIN_MTU; | |
1092 | dev->max_mtu = MAX_MTU - dev->hard_header_len; | |
1da177e4 LT |
1093 | } |
1094 | ||
1095 | /* Character device part */ | |
1096 | ||
1097 | /* Poll */ | |
c8d68e6b | 1098 | static unsigned int tun_chr_poll(struct file *file, poll_table *wait) |
6aa20a22 | 1099 | { |
b2430de3 EB |
1100 | struct tun_file *tfile = file->private_data; |
1101 | struct tun_struct *tun = __tun_get(tfile); | |
3c8a9c63 | 1102 | struct sock *sk; |
33dccbb0 | 1103 | unsigned int mask = 0; |
1da177e4 LT |
1104 | |
1105 | if (!tun) | |
eac9e902 | 1106 | return POLLERR; |
1da177e4 | 1107 | |
54f968d6 | 1108 | sk = tfile->socket.sk; |
3c8a9c63 | 1109 | |
6b8a66ee | 1110 | tun_debug(KERN_INFO, tun, "tun_chr_poll\n"); |
1da177e4 | 1111 | |
9e641bdc | 1112 | poll_wait(file, sk_sleep(sk), wait); |
6aa20a22 | 1113 | |
1576d986 | 1114 | if (!skb_array_empty(&tfile->tx_array)) |
1da177e4 LT |
1115 | mask |= POLLIN | POLLRDNORM; |
1116 | ||
b20e2d54 HFS |
1117 | if (tun->dev->flags & IFF_UP && |
1118 | (sock_writeable(sk) || | |
1119 | (!test_and_set_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags) && | |
1120 | sock_writeable(sk)))) | |
33dccbb0 HX |
1121 | mask |= POLLOUT | POLLWRNORM; |
1122 | ||
c70f1829 EB |
1123 | if (tun->dev->reg_state != NETREG_REGISTERED) |
1124 | mask = POLLERR; | |
1125 | ||
631ab46b | 1126 | tun_put(tun); |
1da177e4 LT |
1127 | return mask; |
1128 | } | |
1129 | ||
f42157cb RR |
1130 | /* prepad is the amount to reserve at front. len is length after that. |
1131 | * linear is a hint as to how much to copy (usually headers). */ | |
54f968d6 | 1132 | static struct sk_buff *tun_alloc_skb(struct tun_file *tfile, |
6f7c156c | 1133 | size_t prepad, size_t len, |
1134 | size_t linear, int noblock) | |
f42157cb | 1135 | { |
54f968d6 | 1136 | struct sock *sk = tfile->socket.sk; |
f42157cb | 1137 | struct sk_buff *skb; |
33dccbb0 | 1138 | int err; |
f42157cb RR |
1139 | |
1140 | /* Under a page? Don't bother with paged skb. */ | |
0eca93bc | 1141 | if (prepad + len < PAGE_SIZE || !linear) |
33dccbb0 | 1142 | linear = len; |
f42157cb | 1143 | |
33dccbb0 | 1144 | skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock, |
28d64271 | 1145 | &err, 0); |
f42157cb | 1146 | if (!skb) |
33dccbb0 | 1147 | return ERR_PTR(err); |
f42157cb RR |
1148 | |
1149 | skb_reserve(skb, prepad); | |
1150 | skb_put(skb, linear); | |
33dccbb0 HX |
1151 | skb->data_len = len - linear; |
1152 | skb->len += len - linear; | |
f42157cb RR |
1153 | |
1154 | return skb; | |
1155 | } | |
1156 | ||
5503fcec JW |
1157 | static void tun_rx_batched(struct tun_struct *tun, struct tun_file *tfile, |
1158 | struct sk_buff *skb, int more) | |
1159 | { | |
1160 | struct sk_buff_head *queue = &tfile->sk.sk_write_queue; | |
1161 | struct sk_buff_head process_queue; | |
1162 | u32 rx_batched = tun->rx_batched; | |
1163 | bool rcv = false; | |
1164 | ||
1165 | if (!rx_batched || (!more && skb_queue_empty(queue))) { | |
1166 | local_bh_disable(); | |
1167 | netif_receive_skb(skb); | |
1168 | local_bh_enable(); | |
1169 | return; | |
1170 | } | |
1171 | ||
1172 | spin_lock(&queue->lock); | |
1173 | if (!more || skb_queue_len(queue) == rx_batched) { | |
1174 | __skb_queue_head_init(&process_queue); | |
1175 | skb_queue_splice_tail_init(queue, &process_queue); | |
1176 | rcv = true; | |
1177 | } else { | |
1178 | __skb_queue_tail(queue, skb); | |
1179 | } | |
1180 | spin_unlock(&queue->lock); | |
1181 | ||
1182 | if (rcv) { | |
1183 | struct sk_buff *nskb; | |
1184 | ||
1185 | local_bh_disable(); | |
1186 | while ((nskb = __skb_dequeue(&process_queue))) | |
1187 | netif_receive_skb(nskb); | |
1188 | netif_receive_skb(skb); | |
1189 | local_bh_enable(); | |
1190 | } | |
1191 | } | |
1192 | ||
1da177e4 | 1193 | /* Get packet from user space buffer */ |
54f968d6 | 1194 | static ssize_t tun_get_user(struct tun_struct *tun, struct tun_file *tfile, |
f5ff53b4 | 1195 | void *msg_control, struct iov_iter *from, |
5503fcec | 1196 | int noblock, bool more) |
1da177e4 | 1197 | { |
09640e63 | 1198 | struct tun_pi pi = { 0, cpu_to_be16(ETH_P_IP) }; |
1da177e4 | 1199 | struct sk_buff *skb; |
f5ff53b4 | 1200 | size_t total_len = iov_iter_count(from); |
eaea34b2 | 1201 | size_t len = total_len, align = tun->align, linear; |
f43798c2 | 1202 | struct virtio_net_hdr gso = { 0 }; |
608b9977 | 1203 | struct tun_pcpu_stats *stats; |
96f8d9ec | 1204 | int good_linear; |
0690899b MT |
1205 | int copylen; |
1206 | bool zerocopy = false; | |
1207 | int err; | |
49974420 | 1208 | u32 rxhash; |
1da177e4 | 1209 | |
1bd4978a ED |
1210 | if (!(tun->dev->flags & IFF_UP)) |
1211 | return -EIO; | |
1212 | ||
40630b82 | 1213 | if (!(tun->flags & IFF_NO_PI)) { |
15718ea0 | 1214 | if (len < sizeof(pi)) |
1da177e4 | 1215 | return -EINVAL; |
15718ea0 | 1216 | len -= sizeof(pi); |
1da177e4 | 1217 | |
cbbd26b8 | 1218 | if (!copy_from_iter_full(&pi, sizeof(pi), from)) |
1da177e4 LT |
1219 | return -EFAULT; |
1220 | } | |
1221 | ||
40630b82 | 1222 | if (tun->flags & IFF_VNET_HDR) { |
e1edab87 WB |
1223 | int vnet_hdr_sz = READ_ONCE(tun->vnet_hdr_sz); |
1224 | ||
1225 | if (len < vnet_hdr_sz) | |
f43798c2 | 1226 | return -EINVAL; |
e1edab87 | 1227 | len -= vnet_hdr_sz; |
f43798c2 | 1228 | |
cbbd26b8 | 1229 | if (!copy_from_iter_full(&gso, sizeof(gso), from)) |
f43798c2 RR |
1230 | return -EFAULT; |
1231 | ||
4909122f | 1232 | if ((gso.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && |
56f0dcc5 MT |
1233 | tun16_to_cpu(tun, gso.csum_start) + tun16_to_cpu(tun, gso.csum_offset) + 2 > tun16_to_cpu(tun, gso.hdr_len)) |
1234 | gso.hdr_len = cpu_to_tun16(tun, tun16_to_cpu(tun, gso.csum_start) + tun16_to_cpu(tun, gso.csum_offset) + 2); | |
4909122f | 1235 | |
56f0dcc5 | 1236 | if (tun16_to_cpu(tun, gso.hdr_len) > len) |
f43798c2 | 1237 | return -EINVAL; |
e1edab87 | 1238 | iov_iter_advance(from, vnet_hdr_sz - sizeof(gso)); |
f43798c2 RR |
1239 | } |
1240 | ||
40630b82 | 1241 | if ((tun->flags & TUN_TYPE_MASK) == IFF_TAP) { |
a504b86e | 1242 | align += NET_IP_ALIGN; |
0eca93bc | 1243 | if (unlikely(len < ETH_HLEN || |
56f0dcc5 | 1244 | (gso.hdr_len && tun16_to_cpu(tun, gso.hdr_len) < ETH_HLEN))) |
e01bf1c8 RR |
1245 | return -EINVAL; |
1246 | } | |
6aa20a22 | 1247 | |
96f8d9ec JW |
1248 | good_linear = SKB_MAX_HEAD(align); |
1249 | ||
88529176 | 1250 | if (msg_control) { |
f5ff53b4 AV |
1251 | struct iov_iter i = *from; |
1252 | ||
88529176 JW |
1253 | /* There are 256 bytes to be copied in skb, so there is |
1254 | * enough room for skb expand head in case it is used. | |
0690899b MT |
1255 | * The rest of the buffer is mapped from userspace. |
1256 | */ | |
56f0dcc5 | 1257 | copylen = gso.hdr_len ? tun16_to_cpu(tun, gso.hdr_len) : GOODCOPY_LEN; |
96f8d9ec JW |
1258 | if (copylen > good_linear) |
1259 | copylen = good_linear; | |
3dd5c330 | 1260 | linear = copylen; |
f5ff53b4 AV |
1261 | iov_iter_advance(&i, copylen); |
1262 | if (iov_iter_npages(&i, INT_MAX) <= MAX_SKB_FRAGS) | |
88529176 JW |
1263 | zerocopy = true; |
1264 | } | |
1265 | ||
1266 | if (!zerocopy) { | |
0690899b | 1267 | copylen = len; |
56f0dcc5 | 1268 | if (tun16_to_cpu(tun, gso.hdr_len) > good_linear) |
96f8d9ec JW |
1269 | linear = good_linear; |
1270 | else | |
56f0dcc5 | 1271 | linear = tun16_to_cpu(tun, gso.hdr_len); |
3dd5c330 | 1272 | } |
0690899b | 1273 | |
3dd5c330 | 1274 | skb = tun_alloc_skb(tfile, align, copylen, linear, noblock); |
33dccbb0 HX |
1275 | if (IS_ERR(skb)) { |
1276 | if (PTR_ERR(skb) != -EAGAIN) | |
608b9977 | 1277 | this_cpu_inc(tun->pcpu_stats->rx_dropped); |
33dccbb0 | 1278 | return PTR_ERR(skb); |
1da177e4 LT |
1279 | } |
1280 | ||
0690899b | 1281 | if (zerocopy) |
f5ff53b4 | 1282 | err = zerocopy_sg_from_iter(skb, from); |
af1cc7a2 | 1283 | else |
f5ff53b4 | 1284 | err = skb_copy_datagram_from_iter(skb, 0, from, len); |
0690899b MT |
1285 | |
1286 | if (err) { | |
608b9977 | 1287 | this_cpu_inc(tun->pcpu_stats->rx_dropped); |
8f22757e | 1288 | kfree_skb(skb); |
1da177e4 | 1289 | return -EFAULT; |
8f22757e | 1290 | } |
1da177e4 | 1291 | |
3e9e40e7 | 1292 | if (virtio_net_hdr_to_skb(skb, &gso, tun_is_little_endian(tun))) { |
df10db98 PA |
1293 | this_cpu_inc(tun->pcpu_stats->rx_frame_errors); |
1294 | kfree_skb(skb); | |
1295 | return -EINVAL; | |
1296 | } | |
1297 | ||
1da177e4 | 1298 | switch (tun->flags & TUN_TYPE_MASK) { |
40630b82 MT |
1299 | case IFF_TUN: |
1300 | if (tun->flags & IFF_NO_PI) { | |
f09f7ee2 AWC |
1301 | switch (skb->data[0] & 0xf0) { |
1302 | case 0x40: | |
1303 | pi.proto = htons(ETH_P_IP); | |
1304 | break; | |
1305 | case 0x60: | |
1306 | pi.proto = htons(ETH_P_IPV6); | |
1307 | break; | |
1308 | default: | |
608b9977 | 1309 | this_cpu_inc(tun->pcpu_stats->rx_dropped); |
f09f7ee2 AWC |
1310 | kfree_skb(skb); |
1311 | return -EINVAL; | |
1312 | } | |
1313 | } | |
1314 | ||
459a98ed | 1315 | skb_reset_mac_header(skb); |
1da177e4 | 1316 | skb->protocol = pi.proto; |
4c13eb66 | 1317 | skb->dev = tun->dev; |
1da177e4 | 1318 | break; |
40630b82 | 1319 | case IFF_TAP: |
1da177e4 LT |
1320 | skb->protocol = eth_type_trans(skb, tun->dev); |
1321 | break; | |
6403eab1 | 1322 | } |
1da177e4 | 1323 | |
0690899b MT |
1324 | /* copy skb_ubuf_info for callback when skb has no error */ |
1325 | if (zerocopy) { | |
1326 | skb_shinfo(skb)->destructor_arg = msg_control; | |
1327 | skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY; | |
c9af6db4 | 1328 | skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG; |
af1cc7a2 JW |
1329 | } else if (msg_control) { |
1330 | struct ubuf_info *uarg = msg_control; | |
1331 | uarg->callback(uarg, false); | |
0690899b MT |
1332 | } |
1333 | ||
72f65107 | 1334 | skb_reset_network_header(skb); |
40893fd0 | 1335 | skb_probe_transport_header(skb, 0); |
38502af7 | 1336 | |
3958afa1 | 1337 | rxhash = skb_get_hash(skb); |
d4aea20d | 1338 | #ifndef CONFIG_4KSTACKS |
5503fcec | 1339 | tun_rx_batched(tun, tfile, skb, more); |
d4aea20d | 1340 | #else |
1da177e4 | 1341 | netif_rx_ni(skb); |
d4aea20d | 1342 | #endif |
6aa20a22 | 1343 | |
608b9977 PA |
1344 | stats = get_cpu_ptr(tun->pcpu_stats); |
1345 | u64_stats_update_begin(&stats->syncp); | |
1346 | stats->rx_packets++; | |
1347 | stats->rx_bytes += len; | |
1348 | u64_stats_update_end(&stats->syncp); | |
1349 | put_cpu_ptr(stats); | |
1da177e4 | 1350 | |
9e85722d | 1351 | tun_flow_update(tun, rxhash, tfile); |
0690899b | 1352 | return total_len; |
6aa20a22 | 1353 | } |
1da177e4 | 1354 | |
f5ff53b4 | 1355 | static ssize_t tun_chr_write_iter(struct kiocb *iocb, struct iov_iter *from) |
1da177e4 | 1356 | { |
33dccbb0 | 1357 | struct file *file = iocb->ki_filp; |
ab46d779 | 1358 | struct tun_struct *tun = tun_get(file); |
54f968d6 | 1359 | struct tun_file *tfile = file->private_data; |
631ab46b | 1360 | ssize_t result; |
1da177e4 LT |
1361 | |
1362 | if (!tun) | |
1363 | return -EBADFD; | |
1364 | ||
5503fcec JW |
1365 | result = tun_get_user(tun, tfile, NULL, from, |
1366 | file->f_flags & O_NONBLOCK, false); | |
631ab46b EB |
1367 | |
1368 | tun_put(tun); | |
1369 | return result; | |
1da177e4 LT |
1370 | } |
1371 | ||
1da177e4 | 1372 | /* Put packet to the user space buffer */ |
6f7c156c | 1373 | static ssize_t tun_put_user(struct tun_struct *tun, |
54f968d6 | 1374 | struct tun_file *tfile, |
6f7c156c | 1375 | struct sk_buff *skb, |
e0b46d0e | 1376 | struct iov_iter *iter) |
1da177e4 LT |
1377 | { |
1378 | struct tun_pi pi = { 0, skb->protocol }; | |
608b9977 | 1379 | struct tun_pcpu_stats *stats; |
e0b46d0e | 1380 | ssize_t total; |
8c847d25 | 1381 | int vlan_offset = 0; |
a8f9bfdf | 1382 | int vlan_hlen = 0; |
2eb783c4 | 1383 | int vnet_hdr_sz = 0; |
a8f9bfdf | 1384 | |
df8a39de | 1385 | if (skb_vlan_tag_present(skb)) |
a8f9bfdf | 1386 | vlan_hlen = VLAN_HLEN; |
1da177e4 | 1387 | |
40630b82 | 1388 | if (tun->flags & IFF_VNET_HDR) |
e1edab87 | 1389 | vnet_hdr_sz = READ_ONCE(tun->vnet_hdr_sz); |
1da177e4 | 1390 | |
e0b46d0e HX |
1391 | total = skb->len + vlan_hlen + vnet_hdr_sz; |
1392 | ||
40630b82 | 1393 | if (!(tun->flags & IFF_NO_PI)) { |
e0b46d0e | 1394 | if (iov_iter_count(iter) < sizeof(pi)) |
1da177e4 LT |
1395 | return -EINVAL; |
1396 | ||
e0b46d0e HX |
1397 | total += sizeof(pi); |
1398 | if (iov_iter_count(iter) < total) { | |
1da177e4 LT |
1399 | /* Packet will be striped */ |
1400 | pi.flags |= TUN_PKT_STRIP; | |
1401 | } | |
6aa20a22 | 1402 | |
e0b46d0e | 1403 | if (copy_to_iter(&pi, sizeof(pi), iter) != sizeof(pi)) |
1da177e4 | 1404 | return -EFAULT; |
6aa20a22 | 1405 | } |
1da177e4 | 1406 | |
2eb783c4 | 1407 | if (vnet_hdr_sz) { |
9403cd7c | 1408 | struct virtio_net_hdr gso; |
34166093 | 1409 | |
e0b46d0e | 1410 | if (iov_iter_count(iter) < vnet_hdr_sz) |
f43798c2 RR |
1411 | return -EINVAL; |
1412 | ||
3e9e40e7 | 1413 | if (virtio_net_hdr_from_skb(skb, &gso, |
6391a448 | 1414 | tun_is_little_endian(tun), true)) { |
f43798c2 | 1415 | struct skb_shared_info *sinfo = skb_shinfo(skb); |
34166093 MR |
1416 | pr_err("unexpected GSO type: " |
1417 | "0x%x, gso_size %d, hdr_len %d\n", | |
1418 | sinfo->gso_type, tun16_to_cpu(tun, gso.gso_size), | |
1419 | tun16_to_cpu(tun, gso.hdr_len)); | |
1420 | print_hex_dump(KERN_ERR, "tun: ", | |
1421 | DUMP_PREFIX_NONE, | |
1422 | 16, 1, skb->head, | |
1423 | min((int)tun16_to_cpu(tun, gso.hdr_len), 64), true); | |
1424 | WARN_ON_ONCE(1); | |
1425 | return -EINVAL; | |
1426 | } | |
f43798c2 | 1427 | |
e0b46d0e | 1428 | if (copy_to_iter(&gso, sizeof(gso), iter) != sizeof(gso)) |
f43798c2 | 1429 | return -EFAULT; |
8c847d25 JW |
1430 | |
1431 | iov_iter_advance(iter, vnet_hdr_sz - sizeof(gso)); | |
f43798c2 RR |
1432 | } |
1433 | ||
a8f9bfdf | 1434 | if (vlan_hlen) { |
e0b46d0e | 1435 | int ret; |
6680ec68 JW |
1436 | struct { |
1437 | __be16 h_vlan_proto; | |
1438 | __be16 h_vlan_TCI; | |
1439 | } veth; | |
1440 | ||
1441 | veth.h_vlan_proto = skb->vlan_proto; | |
df8a39de | 1442 | veth.h_vlan_TCI = htons(skb_vlan_tag_get(skb)); |
6680ec68 JW |
1443 | |
1444 | vlan_offset = offsetof(struct vlan_ethhdr, h_vlan_proto); | |
6680ec68 | 1445 | |
e0b46d0e HX |
1446 | ret = skb_copy_datagram_iter(skb, 0, iter, vlan_offset); |
1447 | if (ret || !iov_iter_count(iter)) | |
6680ec68 JW |
1448 | goto done; |
1449 | ||
e0b46d0e HX |
1450 | ret = copy_to_iter(&veth, sizeof(veth), iter); |
1451 | if (ret != sizeof(veth) || !iov_iter_count(iter)) | |
6680ec68 JW |
1452 | goto done; |
1453 | } | |
1da177e4 | 1454 | |
e0b46d0e | 1455 | skb_copy_datagram_iter(skb, vlan_offset, iter, skb->len - vlan_offset); |
1da177e4 | 1456 | |
6680ec68 | 1457 | done: |
608b9977 PA |
1458 | /* caller is in process context, */ |
1459 | stats = get_cpu_ptr(tun->pcpu_stats); | |
1460 | u64_stats_update_begin(&stats->syncp); | |
1461 | stats->tx_packets++; | |
1462 | stats->tx_bytes += skb->len + vlan_hlen; | |
1463 | u64_stats_update_end(&stats->syncp); | |
1464 | put_cpu_ptr(tun->pcpu_stats); | |
1da177e4 LT |
1465 | |
1466 | return total; | |
1467 | } | |
1468 | ||
1576d986 JW |
1469 | static struct sk_buff *tun_ring_recv(struct tun_file *tfile, int noblock, |
1470 | int *err) | |
1471 | { | |
1472 | DECLARE_WAITQUEUE(wait, current); | |
1473 | struct sk_buff *skb = NULL; | |
f48cc6b2 | 1474 | int error = 0; |
1576d986 JW |
1475 | |
1476 | skb = skb_array_consume(&tfile->tx_array); | |
1477 | if (skb) | |
1478 | goto out; | |
1479 | if (noblock) { | |
f48cc6b2 | 1480 | error = -EAGAIN; |
1576d986 JW |
1481 | goto out; |
1482 | } | |
1483 | ||
1484 | add_wait_queue(&tfile->wq.wait, &wait); | |
1485 | current->state = TASK_INTERRUPTIBLE; | |
1486 | ||
1487 | while (1) { | |
1488 | skb = skb_array_consume(&tfile->tx_array); | |
1489 | if (skb) | |
1490 | break; | |
1491 | if (signal_pending(current)) { | |
f48cc6b2 | 1492 | error = -ERESTARTSYS; |
1576d986 JW |
1493 | break; |
1494 | } | |
1495 | if (tfile->socket.sk->sk_shutdown & RCV_SHUTDOWN) { | |
f48cc6b2 | 1496 | error = -EFAULT; |
1576d986 JW |
1497 | break; |
1498 | } | |
1499 | ||
1500 | schedule(); | |
1501 | } | |
1502 | ||
1503 | current->state = TASK_RUNNING; | |
1504 | remove_wait_queue(&tfile->wq.wait, &wait); | |
1505 | ||
1506 | out: | |
f48cc6b2 | 1507 | *err = error; |
1576d986 JW |
1508 | return skb; |
1509 | } | |
1510 | ||
54f968d6 | 1511 | static ssize_t tun_do_read(struct tun_struct *tun, struct tun_file *tfile, |
9b067034 AV |
1512 | struct iov_iter *to, |
1513 | int noblock) | |
1da177e4 | 1514 | { |
1da177e4 | 1515 | struct sk_buff *skb; |
9b067034 | 1516 | ssize_t ret; |
1576d986 | 1517 | int err; |
1da177e4 | 1518 | |
3872baf6 | 1519 | tun_debug(KERN_INFO, tun, "tun_do_read\n"); |
1da177e4 | 1520 | |
9b067034 AV |
1521 | if (!iov_iter_count(to)) |
1522 | return 0; | |
1da177e4 | 1523 | |
1576d986 JW |
1524 | /* Read frames from ring */ |
1525 | skb = tun_ring_recv(tfile, noblock, &err); | |
e0b46d0e | 1526 | if (!skb) |
957f094f | 1527 | return err; |
e0b46d0e | 1528 | |
9b067034 | 1529 | ret = tun_put_user(tun, tfile, skb, to); |
f51a5e82 | 1530 | if (unlikely(ret < 0)) |
f271b2cc | 1531 | kfree_skb(skb); |
f51a5e82 JW |
1532 | else |
1533 | consume_skb(skb); | |
1da177e4 | 1534 | |
05c2828c MT |
1535 | return ret; |
1536 | } | |
1537 | ||
9b067034 | 1538 | static ssize_t tun_chr_read_iter(struct kiocb *iocb, struct iov_iter *to) |
05c2828c MT |
1539 | { |
1540 | struct file *file = iocb->ki_filp; | |
1541 | struct tun_file *tfile = file->private_data; | |
1542 | struct tun_struct *tun = __tun_get(tfile); | |
9b067034 | 1543 | ssize_t len = iov_iter_count(to), ret; |
05c2828c MT |
1544 | |
1545 | if (!tun) | |
1546 | return -EBADFD; | |
9b067034 | 1547 | ret = tun_do_read(tun, tfile, to, file->f_flags & O_NONBLOCK); |
42404c09 | 1548 | ret = min_t(ssize_t, ret, len); |
d0b7da8a ZYW |
1549 | if (ret > 0) |
1550 | iocb->ki_pos = ret; | |
631ab46b | 1551 | tun_put(tun); |
1da177e4 LT |
1552 | return ret; |
1553 | } | |
1554 | ||
96442e42 JW |
1555 | static void tun_free_netdev(struct net_device *dev) |
1556 | { | |
1557 | struct tun_struct *tun = netdev_priv(dev); | |
1558 | ||
4008e97f | 1559 | BUG_ON(!(list_empty(&tun->disabled))); |
608b9977 | 1560 | free_percpu(tun->pcpu_stats); |
96442e42 | 1561 | tun_flow_uninit(tun); |
5dbbaf2d | 1562 | security_tun_dev_free_security(tun->security); |
96442e42 JW |
1563 | } |
1564 | ||
1da177e4 LT |
1565 | static void tun_setup(struct net_device *dev) |
1566 | { | |
1567 | struct tun_struct *tun = netdev_priv(dev); | |
1568 | ||
0625c883 EB |
1569 | tun->owner = INVALID_UID; |
1570 | tun->group = INVALID_GID; | |
1da177e4 | 1571 | |
1da177e4 | 1572 | dev->ethtool_ops = &tun_ethtool_ops; |
cf124db5 DM |
1573 | dev->needs_free_netdev = true; |
1574 | dev->priv_destructor = tun_free_netdev; | |
016adb72 JW |
1575 | /* We prefer our own queue length */ |
1576 | dev->tx_queue_len = TUN_READQ_SIZE; | |
1da177e4 LT |
1577 | } |
1578 | ||
f019a7a5 EB |
1579 | /* Trivial set of netlink ops to allow deleting tun or tap |
1580 | * device with netlink. | |
1581 | */ | |
1582 | static int tun_validate(struct nlattr *tb[], struct nlattr *data[]) | |
1583 | { | |
1584 | return -EINVAL; | |
1585 | } | |
1586 | ||
1587 | static struct rtnl_link_ops tun_link_ops __read_mostly = { | |
1588 | .kind = DRV_NAME, | |
1589 | .priv_size = sizeof(struct tun_struct), | |
1590 | .setup = tun_setup, | |
1591 | .validate = tun_validate, | |
1592 | }; | |
1593 | ||
33dccbb0 HX |
1594 | static void tun_sock_write_space(struct sock *sk) |
1595 | { | |
54f968d6 | 1596 | struct tun_file *tfile; |
43815482 | 1597 | wait_queue_head_t *wqueue; |
33dccbb0 HX |
1598 | |
1599 | if (!sock_writeable(sk)) | |
1600 | return; | |
1601 | ||
9cd3e072 | 1602 | if (!test_and_clear_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags)) |
33dccbb0 HX |
1603 | return; |
1604 | ||
43815482 ED |
1605 | wqueue = sk_sleep(sk); |
1606 | if (wqueue && waitqueue_active(wqueue)) | |
1607 | wake_up_interruptible_sync_poll(wqueue, POLLOUT | | |
05c2828c | 1608 | POLLWRNORM | POLLWRBAND); |
c722c625 | 1609 | |
54f968d6 JW |
1610 | tfile = container_of(sk, struct tun_file, sk); |
1611 | kill_fasync(&tfile->fasync, SIGIO, POLL_OUT); | |
33dccbb0 HX |
1612 | } |
1613 | ||
1b784140 | 1614 | static int tun_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) |
05c2828c | 1615 | { |
54f968d6 JW |
1616 | int ret; |
1617 | struct tun_file *tfile = container_of(sock, struct tun_file, socket); | |
1618 | struct tun_struct *tun = __tun_get(tfile); | |
1619 | ||
1620 | if (!tun) | |
1621 | return -EBADFD; | |
f5ff53b4 | 1622 | |
c0371da6 | 1623 | ret = tun_get_user(tun, tfile, m->msg_control, &m->msg_iter, |
5503fcec JW |
1624 | m->msg_flags & MSG_DONTWAIT, |
1625 | m->msg_flags & MSG_MORE); | |
54f968d6 JW |
1626 | tun_put(tun); |
1627 | return ret; | |
05c2828c MT |
1628 | } |
1629 | ||
1b784140 | 1630 | static int tun_recvmsg(struct socket *sock, struct msghdr *m, size_t total_len, |
05c2828c MT |
1631 | int flags) |
1632 | { | |
54f968d6 JW |
1633 | struct tun_file *tfile = container_of(sock, struct tun_file, socket); |
1634 | struct tun_struct *tun = __tun_get(tfile); | |
05c2828c | 1635 | int ret; |
54f968d6 JW |
1636 | |
1637 | if (!tun) | |
1638 | return -EBADFD; | |
1639 | ||
eda29772 | 1640 | if (flags & ~(MSG_DONTWAIT|MSG_TRUNC|MSG_ERRQUEUE)) { |
3811ae76 G |
1641 | ret = -EINVAL; |
1642 | goto out; | |
1643 | } | |
eda29772 RC |
1644 | if (flags & MSG_ERRQUEUE) { |
1645 | ret = sock_recv_errqueue(sock->sk, m, total_len, | |
1646 | SOL_PACKET, TUN_TX_TIMESTAMP); | |
1647 | goto out; | |
1648 | } | |
c0371da6 | 1649 | ret = tun_do_read(tun, tfile, &m->msg_iter, flags & MSG_DONTWAIT); |
87897931 | 1650 | if (ret > (ssize_t)total_len) { |
42404c09 DM |
1651 | m->msg_flags |= MSG_TRUNC; |
1652 | ret = flags & MSG_TRUNC ? ret : total_len; | |
1653 | } | |
3811ae76 | 1654 | out: |
54f968d6 | 1655 | tun_put(tun); |
05c2828c MT |
1656 | return ret; |
1657 | } | |
1658 | ||
1576d986 JW |
1659 | static int tun_peek_len(struct socket *sock) |
1660 | { | |
1661 | struct tun_file *tfile = container_of(sock, struct tun_file, socket); | |
1662 | struct tun_struct *tun; | |
1663 | int ret = 0; | |
1664 | ||
1665 | tun = __tun_get(tfile); | |
1666 | if (!tun) | |
1667 | return 0; | |
1668 | ||
1669 | ret = skb_array_peek_len(&tfile->tx_array); | |
1670 | tun_put(tun); | |
1671 | ||
1672 | return ret; | |
1673 | } | |
1674 | ||
05c2828c MT |
1675 | /* Ops structure to mimic raw sockets with tun */ |
1676 | static const struct proto_ops tun_socket_ops = { | |
1576d986 | 1677 | .peek_len = tun_peek_len, |
05c2828c MT |
1678 | .sendmsg = tun_sendmsg, |
1679 | .recvmsg = tun_recvmsg, | |
1680 | }; | |
1681 | ||
33dccbb0 HX |
1682 | static struct proto tun_proto = { |
1683 | .name = "tun", | |
1684 | .owner = THIS_MODULE, | |
54f968d6 | 1685 | .obj_size = sizeof(struct tun_file), |
33dccbb0 | 1686 | }; |
f019a7a5 | 1687 | |
980c9e8c DW |
1688 | static int tun_flags(struct tun_struct *tun) |
1689 | { | |
031f5e03 | 1690 | return tun->flags & (TUN_FEATURES | IFF_PERSIST | IFF_TUN | IFF_TAP); |
980c9e8c DW |
1691 | } |
1692 | ||
1693 | static ssize_t tun_show_flags(struct device *dev, struct device_attribute *attr, | |
1694 | char *buf) | |
1695 | { | |
1696 | struct tun_struct *tun = netdev_priv(to_net_dev(dev)); | |
1697 | return sprintf(buf, "0x%x\n", tun_flags(tun)); | |
1698 | } | |
1699 | ||
1700 | static ssize_t tun_show_owner(struct device *dev, struct device_attribute *attr, | |
1701 | char *buf) | |
1702 | { | |
1703 | struct tun_struct *tun = netdev_priv(to_net_dev(dev)); | |
0625c883 EB |
1704 | return uid_valid(tun->owner)? |
1705 | sprintf(buf, "%u\n", | |
1706 | from_kuid_munged(current_user_ns(), tun->owner)): | |
1707 | sprintf(buf, "-1\n"); | |
980c9e8c DW |
1708 | } |
1709 | ||
1710 | static ssize_t tun_show_group(struct device *dev, struct device_attribute *attr, | |
1711 | char *buf) | |
1712 | { | |
1713 | struct tun_struct *tun = netdev_priv(to_net_dev(dev)); | |
0625c883 EB |
1714 | return gid_valid(tun->group) ? |
1715 | sprintf(buf, "%u\n", | |
1716 | from_kgid_munged(current_user_ns(), tun->group)): | |
1717 | sprintf(buf, "-1\n"); | |
980c9e8c DW |
1718 | } |
1719 | ||
1720 | static DEVICE_ATTR(tun_flags, 0444, tun_show_flags, NULL); | |
1721 | static DEVICE_ATTR(owner, 0444, tun_show_owner, NULL); | |
1722 | static DEVICE_ATTR(group, 0444, tun_show_group, NULL); | |
1723 | ||
c4d33e24 TI |
1724 | static struct attribute *tun_dev_attrs[] = { |
1725 | &dev_attr_tun_flags.attr, | |
1726 | &dev_attr_owner.attr, | |
1727 | &dev_attr_group.attr, | |
1728 | NULL | |
1729 | }; | |
1730 | ||
1731 | static const struct attribute_group tun_attr_group = { | |
1732 | .attrs = tun_dev_attrs | |
1733 | }; | |
1734 | ||
d647a591 | 1735 | static int tun_set_iff(struct net *net, struct file *file, struct ifreq *ifr) |
1da177e4 LT |
1736 | { |
1737 | struct tun_struct *tun; | |
54f968d6 | 1738 | struct tun_file *tfile = file->private_data; |
1da177e4 LT |
1739 | struct net_device *dev; |
1740 | int err; | |
1741 | ||
7c0c3b1a JW |
1742 | if (tfile->detached) |
1743 | return -EINVAL; | |
1744 | ||
74a3e5a7 EB |
1745 | dev = __dev_get_by_name(net, ifr->ifr_name); |
1746 | if (dev) { | |
f85ba780 DW |
1747 | if (ifr->ifr_flags & IFF_TUN_EXCL) |
1748 | return -EBUSY; | |
74a3e5a7 EB |
1749 | if ((ifr->ifr_flags & IFF_TUN) && dev->netdev_ops == &tun_netdev_ops) |
1750 | tun = netdev_priv(dev); | |
1751 | else if ((ifr->ifr_flags & IFF_TAP) && dev->netdev_ops == &tap_netdev_ops) | |
1752 | tun = netdev_priv(dev); | |
1753 | else | |
1754 | return -EINVAL; | |
1755 | ||
8e6d91ae | 1756 | if (!!(ifr->ifr_flags & IFF_MULTI_QUEUE) != |
40630b82 | 1757 | !!(tun->flags & IFF_MULTI_QUEUE)) |
8e6d91ae JW |
1758 | return -EINVAL; |
1759 | ||
cde8b15f | 1760 | if (tun_not_capable(tun)) |
2b980dbd | 1761 | return -EPERM; |
5dbbaf2d | 1762 | err = security_tun_dev_open(tun->security); |
2b980dbd PM |
1763 | if (err < 0) |
1764 | return err; | |
1765 | ||
849c9b6f | 1766 | err = tun_attach(tun, file, ifr->ifr_flags & IFF_NOFILTER); |
a7385ba2 EB |
1767 | if (err < 0) |
1768 | return err; | |
4008e97f | 1769 | |
40630b82 | 1770 | if (tun->flags & IFF_MULTI_QUEUE && |
e8dbad66 JW |
1771 | (tun->numqueues + tun->numdisabled > 1)) { |
1772 | /* One or more queue has already been attached, no need | |
1773 | * to initialize the device again. | |
1774 | */ | |
1775 | return 0; | |
1776 | } | |
6aa20a22 | 1777 | } |
1da177e4 LT |
1778 | else { |
1779 | char *name; | |
1780 | unsigned long flags = 0; | |
edfb6a14 JW |
1781 | int queues = ifr->ifr_flags & IFF_MULTI_QUEUE ? |
1782 | MAX_TAP_QUEUES : 1; | |
1da177e4 | 1783 | |
c260b772 | 1784 | if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) |
ca6bb5d7 | 1785 | return -EPERM; |
2b980dbd PM |
1786 | err = security_tun_dev_create(); |
1787 | if (err < 0) | |
1788 | return err; | |
ca6bb5d7 | 1789 | |
1da177e4 LT |
1790 | /* Set dev type */ |
1791 | if (ifr->ifr_flags & IFF_TUN) { | |
1792 | /* TUN device */ | |
40630b82 | 1793 | flags |= IFF_TUN; |
1da177e4 LT |
1794 | name = "tun%d"; |
1795 | } else if (ifr->ifr_flags & IFF_TAP) { | |
1796 | /* TAP device */ | |
40630b82 | 1797 | flags |= IFF_TAP; |
1da177e4 | 1798 | name = "tap%d"; |
6aa20a22 | 1799 | } else |
36989b90 | 1800 | return -EINVAL; |
6aa20a22 | 1801 | |
1da177e4 LT |
1802 | if (*ifr->ifr_name) |
1803 | name = ifr->ifr_name; | |
1804 | ||
c8d68e6b | 1805 | dev = alloc_netdev_mqs(sizeof(struct tun_struct), name, |
c835a677 TG |
1806 | NET_NAME_UNKNOWN, tun_setup, queues, |
1807 | queues); | |
edfb6a14 | 1808 | |
1da177e4 LT |
1809 | if (!dev) |
1810 | return -ENOMEM; | |
1811 | ||
fc54c658 | 1812 | dev_net_set(dev, net); |
f019a7a5 | 1813 | dev->rtnl_link_ops = &tun_link_ops; |
fb7589a1 | 1814 | dev->ifindex = tfile->ifindex; |
c4d33e24 | 1815 | dev->sysfs_groups[0] = &tun_attr_group; |
758e43b7 | 1816 | |
1da177e4 LT |
1817 | tun = netdev_priv(dev); |
1818 | tun->dev = dev; | |
1819 | tun->flags = flags; | |
f271b2cc | 1820 | tun->txflt.count = 0; |
d9d52b51 | 1821 | tun->vnet_hdr_sz = sizeof(struct virtio_net_hdr); |
33dccbb0 | 1822 | |
eaea34b2 | 1823 | tun->align = NET_SKB_PAD; |
54f968d6 JW |
1824 | tun->filter_attached = false; |
1825 | tun->sndbuf = tfile->socket.sk->sk_sndbuf; | |
5503fcec | 1826 | tun->rx_batched = 0; |
33dccbb0 | 1827 | |
608b9977 PA |
1828 | tun->pcpu_stats = netdev_alloc_pcpu_stats(struct tun_pcpu_stats); |
1829 | if (!tun->pcpu_stats) { | |
1830 | err = -ENOMEM; | |
1831 | goto err_free_dev; | |
1832 | } | |
1833 | ||
96442e42 JW |
1834 | spin_lock_init(&tun->lock); |
1835 | ||
5dbbaf2d PM |
1836 | err = security_tun_dev_alloc_security(&tun->security); |
1837 | if (err < 0) | |
608b9977 | 1838 | goto err_free_stat; |
2b980dbd | 1839 | |
1da177e4 | 1840 | tun_net_init(dev); |
944a1376 | 1841 | tun_flow_init(tun); |
96442e42 | 1842 | |
88255375 | 1843 | dev->hw_features = NETIF_F_SG | NETIF_F_FRAGLIST | |
6680ec68 JW |
1844 | TUN_USER_FEATURES | NETIF_F_HW_VLAN_CTAG_TX | |
1845 | NETIF_F_HW_VLAN_STAG_TX; | |
2a2bbf17 | 1846 | dev->features = dev->hw_features | NETIF_F_LLTX; |
6671b224 FLVC |
1847 | dev->vlan_features = dev->features & |
1848 | ~(NETIF_F_HW_VLAN_CTAG_TX | | |
1849 | NETIF_F_HW_VLAN_STAG_TX); | |
88255375 | 1850 | |
4008e97f | 1851 | INIT_LIST_HEAD(&tun->disabled); |
849c9b6f | 1852 | err = tun_attach(tun, file, false); |
eb0fb363 | 1853 | if (err < 0) |
662ca437 | 1854 | goto err_free_flow; |
eb0fb363 | 1855 | |
1da177e4 LT |
1856 | err = register_netdevice(tun->dev); |
1857 | if (err < 0) | |
662ca437 | 1858 | goto err_detach; |
1da177e4 LT |
1859 | } |
1860 | ||
af668b3c MT |
1861 | netif_carrier_on(tun->dev); |
1862 | ||
6b8a66ee | 1863 | tun_debug(KERN_INFO, tun, "tun_set_iff\n"); |
1da177e4 | 1864 | |
031f5e03 MT |
1865 | tun->flags = (tun->flags & ~TUN_FEATURES) | |
1866 | (ifr->ifr_flags & TUN_FEATURES); | |
c8d68e6b | 1867 | |
e35259a9 MK |
1868 | /* Make sure persistent devices do not get stuck in |
1869 | * xoff state. | |
1870 | */ | |
1871 | if (netif_running(tun->dev)) | |
c8d68e6b | 1872 | netif_tx_wake_all_queues(tun->dev); |
e35259a9 | 1873 | |
1da177e4 LT |
1874 | strcpy(ifr->ifr_name, tun->dev->name); |
1875 | return 0; | |
1876 | ||
662ca437 JW |
1877 | err_detach: |
1878 | tun_detach_all(dev); | |
1879 | err_free_flow: | |
1880 | tun_flow_uninit(tun); | |
1881 | security_tun_dev_free_security(tun->security); | |
608b9977 PA |
1882 | err_free_stat: |
1883 | free_percpu(tun->pcpu_stats); | |
662ca437 | 1884 | err_free_dev: |
1da177e4 | 1885 | free_netdev(dev); |
1da177e4 LT |
1886 | return err; |
1887 | } | |
1888 | ||
9ce99cf6 | 1889 | static void tun_get_iff(struct net *net, struct tun_struct *tun, |
876bfd4d | 1890 | struct ifreq *ifr) |
e3b99556 | 1891 | { |
6b8a66ee | 1892 | tun_debug(KERN_INFO, tun, "tun_get_iff\n"); |
e3b99556 MM |
1893 | |
1894 | strcpy(ifr->ifr_name, tun->dev->name); | |
1895 | ||
980c9e8c | 1896 | ifr->ifr_flags = tun_flags(tun); |
e3b99556 | 1897 | |
e3b99556 MM |
1898 | } |
1899 | ||
5228ddc9 RR |
1900 | /* This is like a cut-down ethtool ops, except done via tun fd so no |
1901 | * privs required. */ | |
88255375 | 1902 | static int set_offload(struct tun_struct *tun, unsigned long arg) |
5228ddc9 | 1903 | { |
c8f44aff | 1904 | netdev_features_t features = 0; |
5228ddc9 RR |
1905 | |
1906 | if (arg & TUN_F_CSUM) { | |
88255375 | 1907 | features |= NETIF_F_HW_CSUM; |
5228ddc9 RR |
1908 | arg &= ~TUN_F_CSUM; |
1909 | ||
1910 | if (arg & (TUN_F_TSO4|TUN_F_TSO6)) { | |
1911 | if (arg & TUN_F_TSO_ECN) { | |
1912 | features |= NETIF_F_TSO_ECN; | |
1913 | arg &= ~TUN_F_TSO_ECN; | |
1914 | } | |
1915 | if (arg & TUN_F_TSO4) | |
1916 | features |= NETIF_F_TSO; | |
1917 | if (arg & TUN_F_TSO6) | |
1918 | features |= NETIF_F_TSO6; | |
1919 | arg &= ~(TUN_F_TSO4|TUN_F_TSO6); | |
1920 | } | |
e3e3c423 VY |
1921 | |
1922 | if (arg & TUN_F_UFO) { | |
1923 | features |= NETIF_F_UFO; | |
1924 | arg &= ~TUN_F_UFO; | |
1925 | } | |
5228ddc9 RR |
1926 | } |
1927 | ||
1928 | /* This gives the user a way to test for new features in future by | |
1929 | * trying to set them. */ | |
1930 | if (arg) | |
1931 | return -EINVAL; | |
1932 | ||
88255375 | 1933 | tun->set_features = features; |
09050957 YI |
1934 | tun->dev->wanted_features &= ~TUN_USER_FEATURES; |
1935 | tun->dev->wanted_features |= features; | |
88255375 | 1936 | netdev_update_features(tun->dev); |
5228ddc9 RR |
1937 | |
1938 | return 0; | |
1939 | } | |
1940 | ||
c8d68e6b JW |
1941 | static void tun_detach_filter(struct tun_struct *tun, int n) |
1942 | { | |
1943 | int i; | |
1944 | struct tun_file *tfile; | |
1945 | ||
1946 | for (i = 0; i < n; i++) { | |
b8deabd3 | 1947 | tfile = rtnl_dereference(tun->tfiles[i]); |
8ced425e HFS |
1948 | lock_sock(tfile->socket.sk); |
1949 | sk_detach_filter(tfile->socket.sk); | |
1950 | release_sock(tfile->socket.sk); | |
c8d68e6b JW |
1951 | } |
1952 | ||
1953 | tun->filter_attached = false; | |
1954 | } | |
1955 | ||
1956 | static int tun_attach_filter(struct tun_struct *tun) | |
1957 | { | |
1958 | int i, ret = 0; | |
1959 | struct tun_file *tfile; | |
1960 | ||
1961 | for (i = 0; i < tun->numqueues; i++) { | |
b8deabd3 | 1962 | tfile = rtnl_dereference(tun->tfiles[i]); |
8ced425e HFS |
1963 | lock_sock(tfile->socket.sk); |
1964 | ret = sk_attach_filter(&tun->fprog, tfile->socket.sk); | |
1965 | release_sock(tfile->socket.sk); | |
c8d68e6b JW |
1966 | if (ret) { |
1967 | tun_detach_filter(tun, i); | |
1968 | return ret; | |
1969 | } | |
1970 | } | |
1971 | ||
1972 | tun->filter_attached = true; | |
1973 | return ret; | |
1974 | } | |
1975 | ||
1976 | static void tun_set_sndbuf(struct tun_struct *tun) | |
1977 | { | |
1978 | struct tun_file *tfile; | |
1979 | int i; | |
1980 | ||
1981 | for (i = 0; i < tun->numqueues; i++) { | |
b8deabd3 | 1982 | tfile = rtnl_dereference(tun->tfiles[i]); |
c8d68e6b JW |
1983 | tfile->socket.sk->sk_sndbuf = tun->sndbuf; |
1984 | } | |
1985 | } | |
1986 | ||
cde8b15f JW |
1987 | static int tun_set_queue(struct file *file, struct ifreq *ifr) |
1988 | { | |
1989 | struct tun_file *tfile = file->private_data; | |
1990 | struct tun_struct *tun; | |
cde8b15f JW |
1991 | int ret = 0; |
1992 | ||
1993 | rtnl_lock(); | |
1994 | ||
1995 | if (ifr->ifr_flags & IFF_ATTACH_QUEUE) { | |
4008e97f | 1996 | tun = tfile->detached; |
5dbbaf2d | 1997 | if (!tun) { |
cde8b15f | 1998 | ret = -EINVAL; |
5dbbaf2d PM |
1999 | goto unlock; |
2000 | } | |
2001 | ret = security_tun_dev_attach_queue(tun->security); | |
2002 | if (ret < 0) | |
2003 | goto unlock; | |
849c9b6f | 2004 | ret = tun_attach(tun, file, false); |
4008e97f | 2005 | } else if (ifr->ifr_flags & IFF_DETACH_QUEUE) { |
b8deabd3 | 2006 | tun = rtnl_dereference(tfile->tun); |
40630b82 | 2007 | if (!tun || !(tun->flags & IFF_MULTI_QUEUE) || tfile->detached) |
4008e97f JW |
2008 | ret = -EINVAL; |
2009 | else | |
2010 | __tun_detach(tfile, false); | |
2011 | } else | |
cde8b15f JW |
2012 | ret = -EINVAL; |
2013 | ||
5dbbaf2d | 2014 | unlock: |
cde8b15f JW |
2015 | rtnl_unlock(); |
2016 | return ret; | |
2017 | } | |
2018 | ||
50857e2a AB |
2019 | static long __tun_chr_ioctl(struct file *file, unsigned int cmd, |
2020 | unsigned long arg, int ifreq_len) | |
1da177e4 | 2021 | { |
36b50bab | 2022 | struct tun_file *tfile = file->private_data; |
631ab46b | 2023 | struct tun_struct *tun; |
1da177e4 LT |
2024 | void __user* argp = (void __user*)arg; |
2025 | struct ifreq ifr; | |
0625c883 EB |
2026 | kuid_t owner; |
2027 | kgid_t group; | |
33dccbb0 | 2028 | int sndbuf; |
d9d52b51 | 2029 | int vnet_hdr_sz; |
fb7589a1 | 2030 | unsigned int ifindex; |
1cf8e410 | 2031 | int le; |
f271b2cc | 2032 | int ret; |
1da177e4 | 2033 | |
20861f26 | 2034 | if (cmd == TUNSETIFF || cmd == TUNSETQUEUE || _IOC_TYPE(cmd) == SOCK_IOC_TYPE) { |
50857e2a | 2035 | if (copy_from_user(&ifr, argp, ifreq_len)) |
1da177e4 | 2036 | return -EFAULT; |
8bbb1813 | 2037 | } else { |
a117dacd | 2038 | memset(&ifr, 0, sizeof(ifr)); |
8bbb1813 | 2039 | } |
631ab46b EB |
2040 | if (cmd == TUNGETFEATURES) { |
2041 | /* Currently this just means: "what IFF flags are valid?". | |
2042 | * This is needed because we never checked for invalid flags on | |
031f5e03 MT |
2043 | * TUNSETIFF. |
2044 | */ | |
2045 | return put_user(IFF_TUN | IFF_TAP | TUN_FEATURES, | |
631ab46b | 2046 | (unsigned int __user*)argp); |
cde8b15f JW |
2047 | } else if (cmd == TUNSETQUEUE) |
2048 | return tun_set_queue(file, &ifr); | |
631ab46b | 2049 | |
c8d68e6b | 2050 | ret = 0; |
876bfd4d HX |
2051 | rtnl_lock(); |
2052 | ||
36b50bab | 2053 | tun = __tun_get(tfile); |
0f16bc13 GF |
2054 | if (cmd == TUNSETIFF) { |
2055 | ret = -EEXIST; | |
2056 | if (tun) | |
2057 | goto unlock; | |
2058 | ||
1da177e4 LT |
2059 | ifr.ifr_name[IFNAMSIZ-1] = '\0'; |
2060 | ||
140e807d | 2061 | ret = tun_set_iff(sock_net(&tfile->sk), file, &ifr); |
1da177e4 | 2062 | |
876bfd4d HX |
2063 | if (ret) |
2064 | goto unlock; | |
1da177e4 | 2065 | |
50857e2a | 2066 | if (copy_to_user(argp, &ifr, ifreq_len)) |
876bfd4d HX |
2067 | ret = -EFAULT; |
2068 | goto unlock; | |
1da177e4 | 2069 | } |
fb7589a1 PE |
2070 | if (cmd == TUNSETIFINDEX) { |
2071 | ret = -EPERM; | |
2072 | if (tun) | |
2073 | goto unlock; | |
2074 | ||
2075 | ret = -EFAULT; | |
2076 | if (copy_from_user(&ifindex, argp, sizeof(ifindex))) | |
2077 | goto unlock; | |
2078 | ||
2079 | ret = 0; | |
2080 | tfile->ifindex = ifindex; | |
2081 | goto unlock; | |
2082 | } | |
1da177e4 | 2083 | |
876bfd4d | 2084 | ret = -EBADFD; |
1da177e4 | 2085 | if (!tun) |
876bfd4d | 2086 | goto unlock; |
1da177e4 | 2087 | |
1e588338 | 2088 | tun_debug(KERN_INFO, tun, "tun_chr_ioctl cmd %u\n", cmd); |
1da177e4 | 2089 | |
631ab46b | 2090 | ret = 0; |
1da177e4 | 2091 | switch (cmd) { |
e3b99556 | 2092 | case TUNGETIFF: |
9ce99cf6 | 2093 | tun_get_iff(current->nsproxy->net_ns, tun, &ifr); |
e3b99556 | 2094 | |
3d407a80 PE |
2095 | if (tfile->detached) |
2096 | ifr.ifr_flags |= IFF_DETACH_QUEUE; | |
849c9b6f PE |
2097 | if (!tfile->socket.sk->sk_filter) |
2098 | ifr.ifr_flags |= IFF_NOFILTER; | |
3d407a80 | 2099 | |
50857e2a | 2100 | if (copy_to_user(argp, &ifr, ifreq_len)) |
631ab46b | 2101 | ret = -EFAULT; |
e3b99556 MM |
2102 | break; |
2103 | ||
1da177e4 LT |
2104 | case TUNSETNOCSUM: |
2105 | /* Disable/Enable checksum */ | |
1da177e4 | 2106 | |
88255375 MM |
2107 | /* [unimplemented] */ |
2108 | tun_debug(KERN_INFO, tun, "ignored: set checksum %s\n", | |
6b8a66ee | 2109 | arg ? "disabled" : "enabled"); |
1da177e4 LT |
2110 | break; |
2111 | ||
2112 | case TUNSETPERSIST: | |
54f968d6 JW |
2113 | /* Disable/Enable persist mode. Keep an extra reference to the |
2114 | * module to prevent the module being unprobed. | |
2115 | */ | |
40630b82 MT |
2116 | if (arg && !(tun->flags & IFF_PERSIST)) { |
2117 | tun->flags |= IFF_PERSIST; | |
54f968d6 | 2118 | __module_get(THIS_MODULE); |
dd38bd85 | 2119 | } |
40630b82 MT |
2120 | if (!arg && (tun->flags & IFF_PERSIST)) { |
2121 | tun->flags &= ~IFF_PERSIST; | |
54f968d6 JW |
2122 | module_put(THIS_MODULE); |
2123 | } | |
1da177e4 | 2124 | |
6b8a66ee JP |
2125 | tun_debug(KERN_INFO, tun, "persist %s\n", |
2126 | arg ? "enabled" : "disabled"); | |
1da177e4 LT |
2127 | break; |
2128 | ||
2129 | case TUNSETOWNER: | |
2130 | /* Set owner of the device */ | |
0625c883 EB |
2131 | owner = make_kuid(current_user_ns(), arg); |
2132 | if (!uid_valid(owner)) { | |
2133 | ret = -EINVAL; | |
2134 | break; | |
2135 | } | |
2136 | tun->owner = owner; | |
1e588338 | 2137 | tun_debug(KERN_INFO, tun, "owner set to %u\n", |
0625c883 | 2138 | from_kuid(&init_user_ns, tun->owner)); |
1da177e4 LT |
2139 | break; |
2140 | ||
8c644623 GG |
2141 | case TUNSETGROUP: |
2142 | /* Set group of the device */ | |
0625c883 EB |
2143 | group = make_kgid(current_user_ns(), arg); |
2144 | if (!gid_valid(group)) { | |
2145 | ret = -EINVAL; | |
2146 | break; | |
2147 | } | |
2148 | tun->group = group; | |
1e588338 | 2149 | tun_debug(KERN_INFO, tun, "group set to %u\n", |
0625c883 | 2150 | from_kgid(&init_user_ns, tun->group)); |
8c644623 GG |
2151 | break; |
2152 | ||
ff4cc3ac MK |
2153 | case TUNSETLINK: |
2154 | /* Only allow setting the type when the interface is down */ | |
2155 | if (tun->dev->flags & IFF_UP) { | |
6b8a66ee JP |
2156 | tun_debug(KERN_INFO, tun, |
2157 | "Linktype set failed because interface is up\n"); | |
48abfe05 | 2158 | ret = -EBUSY; |
ff4cc3ac MK |
2159 | } else { |
2160 | tun->dev->type = (int) arg; | |
6b8a66ee JP |
2161 | tun_debug(KERN_INFO, tun, "linktype set to %d\n", |
2162 | tun->dev->type); | |
48abfe05 | 2163 | ret = 0; |
ff4cc3ac | 2164 | } |
631ab46b | 2165 | break; |
ff4cc3ac | 2166 | |
1da177e4 LT |
2167 | #ifdef TUN_DEBUG |
2168 | case TUNSETDEBUG: | |
2169 | tun->debug = arg; | |
2170 | break; | |
2171 | #endif | |
5228ddc9 | 2172 | case TUNSETOFFLOAD: |
88255375 | 2173 | ret = set_offload(tun, arg); |
631ab46b | 2174 | break; |
5228ddc9 | 2175 | |
f271b2cc MK |
2176 | case TUNSETTXFILTER: |
2177 | /* Can be set only for TAPs */ | |
631ab46b | 2178 | ret = -EINVAL; |
40630b82 | 2179 | if ((tun->flags & TUN_TYPE_MASK) != IFF_TAP) |
631ab46b | 2180 | break; |
c0e5a8c2 | 2181 | ret = update_filter(&tun->txflt, (void __user *)arg); |
631ab46b | 2182 | break; |
1da177e4 LT |
2183 | |
2184 | case SIOCGIFHWADDR: | |
b595076a | 2185 | /* Get hw address */ |
f271b2cc MK |
2186 | memcpy(ifr.ifr_hwaddr.sa_data, tun->dev->dev_addr, ETH_ALEN); |
2187 | ifr.ifr_hwaddr.sa_family = tun->dev->type; | |
50857e2a | 2188 | if (copy_to_user(argp, &ifr, ifreq_len)) |
631ab46b EB |
2189 | ret = -EFAULT; |
2190 | break; | |
1da177e4 LT |
2191 | |
2192 | case SIOCSIFHWADDR: | |
f271b2cc | 2193 | /* Set hw address */ |
6b8a66ee JP |
2194 | tun_debug(KERN_DEBUG, tun, "set hw address: %pM\n", |
2195 | ifr.ifr_hwaddr.sa_data); | |
40102371 | 2196 | |
40102371 | 2197 | ret = dev_set_mac_address(tun->dev, &ifr.ifr_hwaddr); |
631ab46b | 2198 | break; |
33dccbb0 HX |
2199 | |
2200 | case TUNGETSNDBUF: | |
54f968d6 | 2201 | sndbuf = tfile->socket.sk->sk_sndbuf; |
33dccbb0 HX |
2202 | if (copy_to_user(argp, &sndbuf, sizeof(sndbuf))) |
2203 | ret = -EFAULT; | |
2204 | break; | |
2205 | ||
2206 | case TUNSETSNDBUF: | |
2207 | if (copy_from_user(&sndbuf, argp, sizeof(sndbuf))) { | |
2208 | ret = -EFAULT; | |
2209 | break; | |
2210 | } | |
2211 | ||
c8d68e6b JW |
2212 | tun->sndbuf = sndbuf; |
2213 | tun_set_sndbuf(tun); | |
33dccbb0 HX |
2214 | break; |
2215 | ||
d9d52b51 MT |
2216 | case TUNGETVNETHDRSZ: |
2217 | vnet_hdr_sz = tun->vnet_hdr_sz; | |
2218 | if (copy_to_user(argp, &vnet_hdr_sz, sizeof(vnet_hdr_sz))) | |
2219 | ret = -EFAULT; | |
2220 | break; | |
2221 | ||
2222 | case TUNSETVNETHDRSZ: | |
2223 | if (copy_from_user(&vnet_hdr_sz, argp, sizeof(vnet_hdr_sz))) { | |
2224 | ret = -EFAULT; | |
2225 | break; | |
2226 | } | |
2227 | if (vnet_hdr_sz < (int)sizeof(struct virtio_net_hdr)) { | |
2228 | ret = -EINVAL; | |
2229 | break; | |
2230 | } | |
2231 | ||
2232 | tun->vnet_hdr_sz = vnet_hdr_sz; | |
2233 | break; | |
2234 | ||
1cf8e410 MT |
2235 | case TUNGETVNETLE: |
2236 | le = !!(tun->flags & TUN_VNET_LE); | |
2237 | if (put_user(le, (int __user *)argp)) | |
2238 | ret = -EFAULT; | |
2239 | break; | |
2240 | ||
2241 | case TUNSETVNETLE: | |
2242 | if (get_user(le, (int __user *)argp)) { | |
2243 | ret = -EFAULT; | |
2244 | break; | |
2245 | } | |
2246 | if (le) | |
2247 | tun->flags |= TUN_VNET_LE; | |
2248 | else | |
2249 | tun->flags &= ~TUN_VNET_LE; | |
2250 | break; | |
2251 | ||
8b8e658b GK |
2252 | case TUNGETVNETBE: |
2253 | ret = tun_get_vnet_be(tun, argp); | |
2254 | break; | |
2255 | ||
2256 | case TUNSETVNETBE: | |
2257 | ret = tun_set_vnet_be(tun, argp); | |
2258 | break; | |
2259 | ||
99405162 MT |
2260 | case TUNATTACHFILTER: |
2261 | /* Can be set only for TAPs */ | |
2262 | ret = -EINVAL; | |
40630b82 | 2263 | if ((tun->flags & TUN_TYPE_MASK) != IFF_TAP) |
99405162 MT |
2264 | break; |
2265 | ret = -EFAULT; | |
54f968d6 | 2266 | if (copy_from_user(&tun->fprog, argp, sizeof(tun->fprog))) |
99405162 MT |
2267 | break; |
2268 | ||
c8d68e6b | 2269 | ret = tun_attach_filter(tun); |
99405162 MT |
2270 | break; |
2271 | ||
2272 | case TUNDETACHFILTER: | |
2273 | /* Can be set only for TAPs */ | |
2274 | ret = -EINVAL; | |
40630b82 | 2275 | if ((tun->flags & TUN_TYPE_MASK) != IFF_TAP) |
99405162 | 2276 | break; |
c8d68e6b JW |
2277 | ret = 0; |
2278 | tun_detach_filter(tun, tun->numqueues); | |
99405162 MT |
2279 | break; |
2280 | ||
76975e9c PE |
2281 | case TUNGETFILTER: |
2282 | ret = -EINVAL; | |
40630b82 | 2283 | if ((tun->flags & TUN_TYPE_MASK) != IFF_TAP) |
76975e9c PE |
2284 | break; |
2285 | ret = -EFAULT; | |
2286 | if (copy_to_user(argp, &tun->fprog, sizeof(tun->fprog))) | |
2287 | break; | |
2288 | ret = 0; | |
2289 | break; | |
2290 | ||
1da177e4 | 2291 | default: |
631ab46b EB |
2292 | ret = -EINVAL; |
2293 | break; | |
ee289b64 | 2294 | } |
1da177e4 | 2295 | |
876bfd4d HX |
2296 | unlock: |
2297 | rtnl_unlock(); | |
2298 | if (tun) | |
2299 | tun_put(tun); | |
631ab46b | 2300 | return ret; |
1da177e4 LT |
2301 | } |
2302 | ||
50857e2a AB |
2303 | static long tun_chr_ioctl(struct file *file, |
2304 | unsigned int cmd, unsigned long arg) | |
2305 | { | |
2306 | return __tun_chr_ioctl(file, cmd, arg, sizeof (struct ifreq)); | |
2307 | } | |
2308 | ||
2309 | #ifdef CONFIG_COMPAT | |
2310 | static long tun_chr_compat_ioctl(struct file *file, | |
2311 | unsigned int cmd, unsigned long arg) | |
2312 | { | |
2313 | switch (cmd) { | |
2314 | case TUNSETIFF: | |
2315 | case TUNGETIFF: | |
2316 | case TUNSETTXFILTER: | |
2317 | case TUNGETSNDBUF: | |
2318 | case TUNSETSNDBUF: | |
2319 | case SIOCGIFHWADDR: | |
2320 | case SIOCSIFHWADDR: | |
2321 | arg = (unsigned long)compat_ptr(arg); | |
2322 | break; | |
2323 | default: | |
2324 | arg = (compat_ulong_t)arg; | |
2325 | break; | |
2326 | } | |
2327 | ||
2328 | /* | |
2329 | * compat_ifreq is shorter than ifreq, so we must not access beyond | |
2330 | * the end of that structure. All fields that are used in this | |
2331 | * driver are compatible though, we don't need to convert the | |
2332 | * contents. | |
2333 | */ | |
2334 | return __tun_chr_ioctl(file, cmd, arg, sizeof(struct compat_ifreq)); | |
2335 | } | |
2336 | #endif /* CONFIG_COMPAT */ | |
2337 | ||
1da177e4 LT |
2338 | static int tun_chr_fasync(int fd, struct file *file, int on) |
2339 | { | |
54f968d6 | 2340 | struct tun_file *tfile = file->private_data; |
1da177e4 LT |
2341 | int ret; |
2342 | ||
54f968d6 | 2343 | if ((ret = fasync_helper(fd, file, on, &tfile->fasync)) < 0) |
9d319522 | 2344 | goto out; |
6aa20a22 | 2345 | |
1da177e4 | 2346 | if (on) { |
e0b93edd | 2347 | __f_setown(file, task_pid(current), PIDTYPE_PID, 0); |
54f968d6 | 2348 | tfile->flags |= TUN_FASYNC; |
6aa20a22 | 2349 | } else |
54f968d6 | 2350 | tfile->flags &= ~TUN_FASYNC; |
9d319522 JC |
2351 | ret = 0; |
2352 | out: | |
9d319522 | 2353 | return ret; |
1da177e4 LT |
2354 | } |
2355 | ||
2356 | static int tun_chr_open(struct inode *inode, struct file * file) | |
2357 | { | |
140e807d | 2358 | struct net *net = current->nsproxy->net_ns; |
631ab46b | 2359 | struct tun_file *tfile; |
deed49fb | 2360 | |
6b8a66ee | 2361 | DBG1(KERN_INFO, "tunX: tun_chr_open\n"); |
631ab46b | 2362 | |
140e807d | 2363 | tfile = (struct tun_file *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL, |
11aa9c28 | 2364 | &tun_proto, 0); |
631ab46b EB |
2365 | if (!tfile) |
2366 | return -ENOMEM; | |
c956674b | 2367 | RCU_INIT_POINTER(tfile->tun, NULL); |
54f968d6 | 2368 | tfile->flags = 0; |
fb7589a1 | 2369 | tfile->ifindex = 0; |
54f968d6 | 2370 | |
54f968d6 | 2371 | init_waitqueue_head(&tfile->wq.wait); |
9e641bdc | 2372 | RCU_INIT_POINTER(tfile->socket.wq, &tfile->wq); |
54f968d6 JW |
2373 | |
2374 | tfile->socket.file = file; | |
2375 | tfile->socket.ops = &tun_socket_ops; | |
2376 | ||
2377 | sock_init_data(&tfile->socket, &tfile->sk); | |
54f968d6 JW |
2378 | |
2379 | tfile->sk.sk_write_space = tun_sock_write_space; | |
2380 | tfile->sk.sk_sndbuf = INT_MAX; | |
2381 | ||
631ab46b | 2382 | file->private_data = tfile; |
4008e97f | 2383 | INIT_LIST_HEAD(&tfile->next); |
54f968d6 | 2384 | |
19a6afb2 JW |
2385 | sock_set_flag(&tfile->sk, SOCK_ZEROCOPY); |
2386 | ||
1da177e4 LT |
2387 | return 0; |
2388 | } | |
2389 | ||
2390 | static int tun_chr_close(struct inode *inode, struct file *file) | |
2391 | { | |
631ab46b | 2392 | struct tun_file *tfile = file->private_data; |
1da177e4 | 2393 | |
c8d68e6b | 2394 | tun_detach(tfile, true); |
1da177e4 LT |
2395 | |
2396 | return 0; | |
2397 | } | |
2398 | ||
93e14b6d | 2399 | #ifdef CONFIG_PROC_FS |
a3816ab0 | 2400 | static void tun_chr_show_fdinfo(struct seq_file *m, struct file *f) |
93e14b6d MY |
2401 | { |
2402 | struct tun_struct *tun; | |
2403 | struct ifreq ifr; | |
2404 | ||
2405 | memset(&ifr, 0, sizeof(ifr)); | |
2406 | ||
2407 | rtnl_lock(); | |
2408 | tun = tun_get(f); | |
2409 | if (tun) | |
2410 | tun_get_iff(current->nsproxy->net_ns, tun, &ifr); | |
2411 | rtnl_unlock(); | |
2412 | ||
2413 | if (tun) | |
2414 | tun_put(tun); | |
2415 | ||
a3816ab0 | 2416 | seq_printf(m, "iff:\t%s\n", ifr.ifr_name); |
93e14b6d MY |
2417 | } |
2418 | #endif | |
2419 | ||
d54b1fdb | 2420 | static const struct file_operations tun_fops = { |
6aa20a22 | 2421 | .owner = THIS_MODULE, |
1da177e4 | 2422 | .llseek = no_llseek, |
9b067034 | 2423 | .read_iter = tun_chr_read_iter, |
f5ff53b4 | 2424 | .write_iter = tun_chr_write_iter, |
1da177e4 | 2425 | .poll = tun_chr_poll, |
50857e2a AB |
2426 | .unlocked_ioctl = tun_chr_ioctl, |
2427 | #ifdef CONFIG_COMPAT | |
2428 | .compat_ioctl = tun_chr_compat_ioctl, | |
2429 | #endif | |
1da177e4 LT |
2430 | .open = tun_chr_open, |
2431 | .release = tun_chr_close, | |
93e14b6d MY |
2432 | .fasync = tun_chr_fasync, |
2433 | #ifdef CONFIG_PROC_FS | |
2434 | .show_fdinfo = tun_chr_show_fdinfo, | |
2435 | #endif | |
1da177e4 LT |
2436 | }; |
2437 | ||
2438 | static struct miscdevice tun_miscdev = { | |
2439 | .minor = TUN_MINOR, | |
2440 | .name = "tun", | |
e454cea2 | 2441 | .nodename = "net/tun", |
1da177e4 | 2442 | .fops = &tun_fops, |
1da177e4 LT |
2443 | }; |
2444 | ||
2445 | /* ethtool interface */ | |
2446 | ||
29ccc49d PR |
2447 | static int tun_get_link_ksettings(struct net_device *dev, |
2448 | struct ethtool_link_ksettings *cmd) | |
2449 | { | |
2450 | ethtool_link_ksettings_zero_link_mode(cmd, supported); | |
2451 | ethtool_link_ksettings_zero_link_mode(cmd, advertising); | |
2452 | cmd->base.speed = SPEED_10; | |
2453 | cmd->base.duplex = DUPLEX_FULL; | |
2454 | cmd->base.port = PORT_TP; | |
2455 | cmd->base.phy_address = 0; | |
2456 | cmd->base.autoneg = AUTONEG_DISABLE; | |
1da177e4 LT |
2457 | return 0; |
2458 | } | |
2459 | ||
2460 | static void tun_get_drvinfo(struct net_device *dev, struct ethtool_drvinfo *info) | |
2461 | { | |
2462 | struct tun_struct *tun = netdev_priv(dev); | |
2463 | ||
33a5ba14 RJ |
2464 | strlcpy(info->driver, DRV_NAME, sizeof(info->driver)); |
2465 | strlcpy(info->version, DRV_VERSION, sizeof(info->version)); | |
1da177e4 LT |
2466 | |
2467 | switch (tun->flags & TUN_TYPE_MASK) { | |
40630b82 | 2468 | case IFF_TUN: |
33a5ba14 | 2469 | strlcpy(info->bus_info, "tun", sizeof(info->bus_info)); |
1da177e4 | 2470 | break; |
40630b82 | 2471 | case IFF_TAP: |
33a5ba14 | 2472 | strlcpy(info->bus_info, "tap", sizeof(info->bus_info)); |
1da177e4 LT |
2473 | break; |
2474 | } | |
2475 | } | |
2476 | ||
2477 | static u32 tun_get_msglevel(struct net_device *dev) | |
2478 | { | |
2479 | #ifdef TUN_DEBUG | |
2480 | struct tun_struct *tun = netdev_priv(dev); | |
2481 | return tun->debug; | |
2482 | #else | |
2483 | return -EOPNOTSUPP; | |
2484 | #endif | |
2485 | } | |
2486 | ||
2487 | static void tun_set_msglevel(struct net_device *dev, u32 value) | |
2488 | { | |
2489 | #ifdef TUN_DEBUG | |
2490 | struct tun_struct *tun = netdev_priv(dev); | |
2491 | tun->debug = value; | |
2492 | #endif | |
2493 | } | |
2494 | ||
5503fcec JW |
2495 | static int tun_get_coalesce(struct net_device *dev, |
2496 | struct ethtool_coalesce *ec) | |
2497 | { | |
2498 | struct tun_struct *tun = netdev_priv(dev); | |
2499 | ||
2500 | ec->rx_max_coalesced_frames = tun->rx_batched; | |
2501 | ||
2502 | return 0; | |
2503 | } | |
2504 | ||
2505 | static int tun_set_coalesce(struct net_device *dev, | |
2506 | struct ethtool_coalesce *ec) | |
2507 | { | |
2508 | struct tun_struct *tun = netdev_priv(dev); | |
2509 | ||
2510 | if (ec->rx_max_coalesced_frames > NAPI_POLL_WEIGHT) | |
2511 | tun->rx_batched = NAPI_POLL_WEIGHT; | |
2512 | else | |
2513 | tun->rx_batched = ec->rx_max_coalesced_frames; | |
2514 | ||
2515 | return 0; | |
2516 | } | |
2517 | ||
7282d491 | 2518 | static const struct ethtool_ops tun_ethtool_ops = { |
1da177e4 LT |
2519 | .get_drvinfo = tun_get_drvinfo, |
2520 | .get_msglevel = tun_get_msglevel, | |
2521 | .set_msglevel = tun_set_msglevel, | |
bee31369 | 2522 | .get_link = ethtool_op_get_link, |
eda29772 | 2523 | .get_ts_info = ethtool_op_get_ts_info, |
5503fcec JW |
2524 | .get_coalesce = tun_get_coalesce, |
2525 | .set_coalesce = tun_set_coalesce, | |
29ccc49d | 2526 | .get_link_ksettings = tun_get_link_ksettings, |
1da177e4 LT |
2527 | }; |
2528 | ||
1576d986 JW |
2529 | static int tun_queue_resize(struct tun_struct *tun) |
2530 | { | |
2531 | struct net_device *dev = tun->dev; | |
2532 | struct tun_file *tfile; | |
2533 | struct skb_array **arrays; | |
2534 | int n = tun->numqueues + tun->numdisabled; | |
2535 | int ret, i; | |
2536 | ||
2537 | arrays = kmalloc(sizeof *arrays * n, GFP_KERNEL); | |
2538 | if (!arrays) | |
2539 | return -ENOMEM; | |
2540 | ||
2541 | for (i = 0; i < tun->numqueues; i++) { | |
2542 | tfile = rtnl_dereference(tun->tfiles[i]); | |
2543 | arrays[i] = &tfile->tx_array; | |
2544 | } | |
2545 | list_for_each_entry(tfile, &tun->disabled, next) | |
2546 | arrays[i++] = &tfile->tx_array; | |
2547 | ||
2548 | ret = skb_array_resize_multiple(arrays, n, | |
2549 | dev->tx_queue_len, GFP_KERNEL); | |
2550 | ||
2551 | kfree(arrays); | |
2552 | return ret; | |
2553 | } | |
2554 | ||
2555 | static int tun_device_event(struct notifier_block *unused, | |
2556 | unsigned long event, void *ptr) | |
2557 | { | |
2558 | struct net_device *dev = netdev_notifier_info_to_dev(ptr); | |
2559 | struct tun_struct *tun = netdev_priv(dev); | |
2560 | ||
86dfb4ac CG |
2561 | if (dev->rtnl_link_ops != &tun_link_ops) |
2562 | return NOTIFY_DONE; | |
2563 | ||
1576d986 JW |
2564 | switch (event) { |
2565 | case NETDEV_CHANGE_TX_QUEUE_LEN: | |
2566 | if (tun_queue_resize(tun)) | |
2567 | return NOTIFY_BAD; | |
2568 | break; | |
2569 | default: | |
2570 | break; | |
2571 | } | |
2572 | ||
2573 | return NOTIFY_DONE; | |
2574 | } | |
2575 | ||
2576 | static struct notifier_block tun_notifier_block __read_mostly = { | |
2577 | .notifier_call = tun_device_event, | |
2578 | }; | |
79d17604 | 2579 | |
1da177e4 LT |
2580 | static int __init tun_init(void) |
2581 | { | |
2582 | int ret = 0; | |
2583 | ||
6b8a66ee | 2584 | pr_info("%s, %s\n", DRV_DESCRIPTION, DRV_VERSION); |
1da177e4 | 2585 | |
f019a7a5 | 2586 | ret = rtnl_link_register(&tun_link_ops); |
79d17604 | 2587 | if (ret) { |
6b8a66ee | 2588 | pr_err("Can't register link_ops\n"); |
f019a7a5 | 2589 | goto err_linkops; |
79d17604 PE |
2590 | } |
2591 | ||
1da177e4 | 2592 | ret = misc_register(&tun_miscdev); |
79d17604 | 2593 | if (ret) { |
6b8a66ee | 2594 | pr_err("Can't register misc device %d\n", TUN_MINOR); |
79d17604 PE |
2595 | goto err_misc; |
2596 | } | |
1576d986 JW |
2597 | |
2598 | register_netdevice_notifier(&tun_notifier_block); | |
f019a7a5 | 2599 | return 0; |
79d17604 | 2600 | err_misc: |
f019a7a5 EB |
2601 | rtnl_link_unregister(&tun_link_ops); |
2602 | err_linkops: | |
1da177e4 LT |
2603 | return ret; |
2604 | } | |
2605 | ||
2606 | static void tun_cleanup(void) | |
2607 | { | |
6aa20a22 | 2608 | misc_deregister(&tun_miscdev); |
f019a7a5 | 2609 | rtnl_link_unregister(&tun_link_ops); |
1576d986 | 2610 | unregister_netdevice_notifier(&tun_notifier_block); |
1da177e4 LT |
2611 | } |
2612 | ||
05c2828c MT |
2613 | /* Get an underlying socket object from tun file. Returns error unless file is |
2614 | * attached to a device. The returned object works like a packet socket, it | |
2615 | * can be used for sock_sendmsg/sock_recvmsg. The caller is responsible for | |
2616 | * holding a reference to the file for as long as the socket is in use. */ | |
2617 | struct socket *tun_get_socket(struct file *file) | |
2618 | { | |
6e914fc7 | 2619 | struct tun_file *tfile; |
05c2828c MT |
2620 | if (file->f_op != &tun_fops) |
2621 | return ERR_PTR(-EINVAL); | |
6e914fc7 JW |
2622 | tfile = file->private_data; |
2623 | if (!tfile) | |
05c2828c | 2624 | return ERR_PTR(-EBADFD); |
54f968d6 | 2625 | return &tfile->socket; |
05c2828c MT |
2626 | } |
2627 | EXPORT_SYMBOL_GPL(tun_get_socket); | |
2628 | ||
1da177e4 LT |
2629 | module_init(tun_init); |
2630 | module_exit(tun_cleanup); | |
2631 | MODULE_DESCRIPTION(DRV_DESCRIPTION); | |
2632 | MODULE_AUTHOR(DRV_COPYRIGHT); | |
2633 | MODULE_LICENSE("GPL"); | |
2634 | MODULE_ALIAS_MISCDEV(TUN_MINOR); | |
578454ff | 2635 | MODULE_ALIAS("devname:net/tun"); |