blob: 2cbf380377d5e009fa972d85af18ded9971a5248 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * NETLINK Kernel-user communication protocol.
3 *
Alan Cox113aa832008-10-13 19:01:08 -07004 * Authors: Alan Cox <alan@lxorguk.ukuu.org.uk>
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +090011 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Tue Jun 26 14:36:48 MEST 2001 Herbert "herp" Rosmanith
13 * added netlink_proto_exit
14 * Tue Jan 22 18:32:44 BRST 2002 Arnaldo C. de Melo <acme@conectiva.com.br>
15 * use nlk_sk, as sk->protinfo is on a diet 8)
Harald Welte4fdb3bb2005-08-09 19:40:55 -070016 * Fri Jul 22 19:51:12 MEST 2005 Harald Welte <laforge@gnumonks.org>
17 * - inc module use count of module that owns
18 * the kernel socket in case userspace opens
19 * socket of same protocol
20 * - remove all module support, since netlink is
21 * mandatory if CONFIG_NET=y these days
Linus Torvalds1da177e2005-04-16 15:20:36 -070022 */
23
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/module.h>
25
Randy Dunlap4fc268d2006-01-11 12:17:47 -080026#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070027#include <linux/kernel.h>
28#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/signal.h>
30#include <linux/sched.h>
31#include <linux/errno.h>
32#include <linux/string.h>
33#include <linux/stat.h>
34#include <linux/socket.h>
35#include <linux/un.h>
36#include <linux/fcntl.h>
37#include <linux/termios.h>
38#include <linux/sockios.h>
39#include <linux/net.h>
40#include <linux/fs.h>
41#include <linux/slab.h>
42#include <asm/uaccess.h>
43#include <linux/skbuff.h>
44#include <linux/netdevice.h>
45#include <linux/rtnetlink.h>
46#include <linux/proc_fs.h>
47#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048#include <linux/notifier.h>
49#include <linux/security.h>
50#include <linux/jhash.h>
51#include <linux/jiffies.h>
52#include <linux/random.h>
53#include <linux/bitops.h>
54#include <linux/mm.h>
55#include <linux/types.h>
Andrew Morton54e0f522005-04-30 07:07:04 +010056#include <linux/audit.h>
Patrick McHardyaf65bdf2007-04-20 14:14:21 -070057#include <linux/mutex.h>
Andrew Morton54e0f522005-04-30 07:07:04 +010058
Eric W. Biederman457c4cb2007-09-12 12:01:34 +020059#include <net/net_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060#include <net/sock.h>
61#include <net/scm.h>
Thomas Graf82ace472005-11-10 02:25:53 +010062#include <net/netlink.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
Patrick McHardyf7fa9b12005-08-15 12:29:13 -070064#define NLGRPSZ(x) (ALIGN(x, sizeof(unsigned long) * 8) / 8)
Johannes Bergb4ff4f02007-07-18 15:46:06 -070065#define NLGRPLONGS(x) (NLGRPSZ(x)/sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
67struct netlink_sock {
68 /* struct sock has to be the first member of netlink_sock */
69 struct sock sk;
70 u32 pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 u32 dst_pid;
Patrick McHardyd629b832005-08-14 19:27:50 -070072 u32 dst_group;
Patrick McHardyf7fa9b12005-08-15 12:29:13 -070073 u32 flags;
74 u32 subscriptions;
75 u32 ngroups;
76 unsigned long *groups;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077 unsigned long state;
78 wait_queue_head_t wait;
79 struct netlink_callback *cb;
Patrick McHardyaf65bdf2007-04-20 14:14:21 -070080 struct mutex *cb_mutex;
81 struct mutex cb_def_mutex;
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -070082 void (*netlink_rcv)(struct sk_buff *skb);
Patrick McHardy77247bb2005-08-14 19:27:13 -070083 struct module *module;
Linus Torvalds1da177e2005-04-16 15:20:36 -070084};
85
Johannes Berg6c04bb12009-07-10 09:51:32 +000086struct listeners_rcu_head {
87 struct rcu_head rcu_head;
88 void *ptr;
89};
90
Patrick McHardy77247bb2005-08-14 19:27:13 -070091#define NETLINK_KERNEL_SOCKET 0x1
Patrick McHardy9a4595b2005-08-15 12:32:15 -070092#define NETLINK_RECV_PKTINFO 0x2
Pablo Neira Ayusobe0c22a2009-02-18 01:40:43 +000093#define NETLINK_BROADCAST_SEND_ERROR 0x4
Pablo Neira Ayuso38938bf2009-03-24 16:37:55 -070094#define NETLINK_RECV_NO_ENOBUFS 0x8
Patrick McHardy77247bb2005-08-14 19:27:13 -070095
Linus Torvalds1da177e2005-04-16 15:20:36 -070096static inline struct netlink_sock *nlk_sk(struct sock *sk)
97{
Denis Cheng32b21e02007-08-28 15:41:11 -070098 return container_of(sk, struct netlink_sock, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -070099}
100
Denis V. Lunevaed81562007-10-10 21:14:32 -0700101static inline int netlink_is_kernel(struct sock *sk)
102{
103 return nlk_sk(sk)->flags & NETLINK_KERNEL_SOCKET;
104}
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106struct nl_pid_hash {
107 struct hlist_head *table;
108 unsigned long rehash_time;
109
110 unsigned int mask;
111 unsigned int shift;
112
113 unsigned int entries;
114 unsigned int max_shift;
115
116 u32 rnd;
117};
118
119struct netlink_table {
120 struct nl_pid_hash hash;
121 struct hlist_head mc_list;
Patrick McHardy4277a082006-03-20 18:52:01 -0800122 unsigned long *listeners;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 unsigned int nl_nonroot;
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700124 unsigned int groups;
Patrick McHardyaf65bdf2007-04-20 14:14:21 -0700125 struct mutex *cb_mutex;
Patrick McHardy77247bb2005-08-14 19:27:13 -0700126 struct module *module;
Patrick McHardyab33a172005-08-14 19:31:36 -0700127 int registered;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128};
129
130static struct netlink_table *nl_table;
131
132static DECLARE_WAIT_QUEUE_HEAD(nl_table_wait);
133
134static int netlink_dump(struct sock *sk);
135static void netlink_destroy_callback(struct netlink_callback *cb);
136
137static DEFINE_RWLOCK(nl_table_lock);
138static atomic_t nl_table_users = ATOMIC_INIT(0);
139
Alan Sterne041c682006-03-27 01:16:30 -0800140static ATOMIC_NOTIFIER_HEAD(netlink_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
Patrick McHardyd629b832005-08-14 19:27:50 -0700142static u32 netlink_group_mask(u32 group)
143{
144 return group ? 1 << (group - 1) : 0;
145}
146
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147static struct hlist_head *nl_pid_hashfn(struct nl_pid_hash *hash, u32 pid)
148{
149 return &hash->table[jhash_1word(pid, hash->rnd) & hash->mask];
150}
151
152static void netlink_sock_destruct(struct sock *sk)
153{
Herbert Xu3f660d62007-05-03 03:17:14 -0700154 struct netlink_sock *nlk = nlk_sk(sk);
155
Herbert Xu3f660d62007-05-03 03:17:14 -0700156 if (nlk->cb) {
157 if (nlk->cb->done)
158 nlk->cb->done(nlk->cb);
159 netlink_destroy_callback(nlk->cb);
160 }
161
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 skb_queue_purge(&sk->sk_receive_queue);
163
164 if (!sock_flag(sk, SOCK_DEAD)) {
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800165 printk(KERN_ERR "Freeing alive netlink socket %p\n", sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 return;
167 }
Ilpo Järvinen547b7922008-07-25 21:43:18 -0700168
169 WARN_ON(atomic_read(&sk->sk_rmem_alloc));
170 WARN_ON(atomic_read(&sk->sk_wmem_alloc));
171 WARN_ON(nlk_sk(sk)->groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172}
173
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800174/* This lock without WQ_FLAG_EXCLUSIVE is good on UP and it is _very_ bad on
175 * SMP. Look, when several writers sleep and reader wakes them up, all but one
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 * immediately hit write lock and grab all the cpus. Exclusive sleep solves
177 * this, _but_ remember, it adds useless work on UP machines.
178 */
179
Johannes Bergd136f1b2009-09-12 03:03:15 +0000180void netlink_table_grab(void)
Eric Dumazet9a429c42008-01-01 21:58:02 -0800181 __acquires(nl_table_lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182{
Johannes Bergd136f1b2009-09-12 03:03:15 +0000183 might_sleep();
184
Arjan van de Ven6abd2192006-07-03 00:24:07 -0700185 write_lock_irq(&nl_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186
187 if (atomic_read(&nl_table_users)) {
188 DECLARE_WAITQUEUE(wait, current);
189
190 add_wait_queue_exclusive(&nl_table_wait, &wait);
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800191 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 set_current_state(TASK_UNINTERRUPTIBLE);
193 if (atomic_read(&nl_table_users) == 0)
194 break;
Arjan van de Ven6abd2192006-07-03 00:24:07 -0700195 write_unlock_irq(&nl_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196 schedule();
Arjan van de Ven6abd2192006-07-03 00:24:07 -0700197 write_lock_irq(&nl_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 }
199
200 __set_current_state(TASK_RUNNING);
201 remove_wait_queue(&nl_table_wait, &wait);
202 }
203}
204
Johannes Bergd136f1b2009-09-12 03:03:15 +0000205void netlink_table_ungrab(void)
Eric Dumazet9a429c42008-01-01 21:58:02 -0800206 __releases(nl_table_lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207{
Arjan van de Ven6abd2192006-07-03 00:24:07 -0700208 write_unlock_irq(&nl_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 wake_up(&nl_table_wait);
210}
211
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800212static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213netlink_lock_table(void)
214{
215 /* read_lock() synchronizes us to netlink_table_grab */
216
217 read_lock(&nl_table_lock);
218 atomic_inc(&nl_table_users);
219 read_unlock(&nl_table_lock);
220}
221
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800222static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223netlink_unlock_table(void)
224{
225 if (atomic_dec_and_test(&nl_table_users))
226 wake_up(&nl_table_wait);
227}
228
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800229static inline struct sock *netlink_lookup(struct net *net, int protocol,
230 u32 pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231{
232 struct nl_pid_hash *hash = &nl_table[protocol].hash;
233 struct hlist_head *head;
234 struct sock *sk;
235 struct hlist_node *node;
236
237 read_lock(&nl_table_lock);
238 head = nl_pid_hashfn(hash, pid);
239 sk_for_each(sk, node, head) {
YOSHIFUJI Hideaki878628f2008-03-26 03:57:35 +0900240 if (net_eq(sock_net(sk), net) && (nlk_sk(sk)->pid == pid)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 sock_hold(sk);
242 goto found;
243 }
244 }
245 sk = NULL;
246found:
247 read_unlock(&nl_table_lock);
248 return sk;
249}
250
Eric Dumazetea729122007-12-11 02:09:47 -0800251static inline struct hlist_head *nl_pid_hash_zalloc(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252{
253 if (size <= PAGE_SIZE)
Eric Dumazetea729122007-12-11 02:09:47 -0800254 return kzalloc(size, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 else
256 return (struct hlist_head *)
Eric Dumazetea729122007-12-11 02:09:47 -0800257 __get_free_pages(GFP_ATOMIC | __GFP_ZERO,
258 get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259}
260
261static inline void nl_pid_hash_free(struct hlist_head *table, size_t size)
262{
263 if (size <= PAGE_SIZE)
264 kfree(table);
265 else
266 free_pages((unsigned long)table, get_order(size));
267}
268
269static int nl_pid_hash_rehash(struct nl_pid_hash *hash, int grow)
270{
271 unsigned int omask, mask, shift;
272 size_t osize, size;
273 struct hlist_head *otable, *table;
274 int i;
275
276 omask = mask = hash->mask;
277 osize = size = (mask + 1) * sizeof(*table);
278 shift = hash->shift;
279
280 if (grow) {
281 if (++shift > hash->max_shift)
282 return 0;
283 mask = mask * 2 + 1;
284 size *= 2;
285 }
286
Eric Dumazetea729122007-12-11 02:09:47 -0800287 table = nl_pid_hash_zalloc(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 if (!table)
289 return 0;
290
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 otable = hash->table;
292 hash->table = table;
293 hash->mask = mask;
294 hash->shift = shift;
295 get_random_bytes(&hash->rnd, sizeof(hash->rnd));
296
297 for (i = 0; i <= omask; i++) {
298 struct sock *sk;
299 struct hlist_node *node, *tmp;
300
301 sk_for_each_safe(sk, node, tmp, &otable[i])
302 __sk_add_node(sk, nl_pid_hashfn(hash, nlk_sk(sk)->pid));
303 }
304
305 nl_pid_hash_free(otable, osize);
306 hash->rehash_time = jiffies + 10 * 60 * HZ;
307 return 1;
308}
309
310static inline int nl_pid_hash_dilute(struct nl_pid_hash *hash, int len)
311{
312 int avg = hash->entries >> hash->shift;
313
314 if (unlikely(avg > 1) && nl_pid_hash_rehash(hash, 1))
315 return 1;
316
317 if (unlikely(len > avg) && time_after(jiffies, hash->rehash_time)) {
318 nl_pid_hash_rehash(hash, 0);
319 return 1;
320 }
321
322 return 0;
323}
324
Eric Dumazet90ddc4f2005-12-22 12:49:22 -0800325static const struct proto_ops netlink_ops;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326
Patrick McHardy4277a082006-03-20 18:52:01 -0800327static void
328netlink_update_listeners(struct sock *sk)
329{
330 struct netlink_table *tbl = &nl_table[sk->sk_protocol];
331 struct hlist_node *node;
332 unsigned long mask;
333 unsigned int i;
334
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700335 for (i = 0; i < NLGRPLONGS(tbl->groups); i++) {
Patrick McHardy4277a082006-03-20 18:52:01 -0800336 mask = 0;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700337 sk_for_each_bound(sk, node, &tbl->mc_list) {
338 if (i < NLGRPLONGS(nlk_sk(sk)->ngroups))
339 mask |= nlk_sk(sk)->groups[i];
340 }
Patrick McHardy4277a082006-03-20 18:52:01 -0800341 tbl->listeners[i] = mask;
342 }
343 /* this function is only called with the netlink table "grabbed", which
344 * makes sure updates are visible before bind or setsockopt return. */
345}
346
Eric W. Biedermanb4b51022007-09-12 13:05:38 +0200347static int netlink_insert(struct sock *sk, struct net *net, u32 pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348{
349 struct nl_pid_hash *hash = &nl_table[sk->sk_protocol].hash;
350 struct hlist_head *head;
351 int err = -EADDRINUSE;
352 struct sock *osk;
353 struct hlist_node *node;
354 int len;
355
356 netlink_table_grab();
357 head = nl_pid_hashfn(hash, pid);
358 len = 0;
359 sk_for_each(osk, node, head) {
YOSHIFUJI Hideaki878628f2008-03-26 03:57:35 +0900360 if (net_eq(sock_net(osk), net) && (nlk_sk(osk)->pid == pid))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 break;
362 len++;
363 }
364 if (node)
365 goto err;
366
367 err = -EBUSY;
368 if (nlk_sk(sk)->pid)
369 goto err;
370
371 err = -ENOMEM;
372 if (BITS_PER_LONG > 32 && unlikely(hash->entries >= UINT_MAX))
373 goto err;
374
375 if (len && nl_pid_hash_dilute(hash, len))
376 head = nl_pid_hashfn(hash, pid);
377 hash->entries++;
378 nlk_sk(sk)->pid = pid;
379 sk_add_node(sk, head);
380 err = 0;
381
382err:
383 netlink_table_ungrab();
384 return err;
385}
386
387static void netlink_remove(struct sock *sk)
388{
389 netlink_table_grab();
David S. Millerd470e3b2005-06-26 15:31:51 -0700390 if (sk_del_node_init(sk))
391 nl_table[sk->sk_protocol].hash.entries--;
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700392 if (nlk_sk(sk)->subscriptions)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 __sk_del_bind_node(sk);
394 netlink_table_ungrab();
395}
396
397static struct proto netlink_proto = {
398 .name = "NETLINK",
399 .owner = THIS_MODULE,
400 .obj_size = sizeof(struct netlink_sock),
401};
402
Eric W. Biederman1b8d7ae2007-10-08 23:24:22 -0700403static int __netlink_create(struct net *net, struct socket *sock,
404 struct mutex *cb_mutex, int protocol)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405{
406 struct sock *sk;
407 struct netlink_sock *nlk;
Patrick McHardyab33a172005-08-14 19:31:36 -0700408
409 sock->ops = &netlink_ops;
410
Pavel Emelyanov6257ff22007-11-01 00:39:31 -0700411 sk = sk_alloc(net, PF_NETLINK, GFP_KERNEL, &netlink_proto);
Patrick McHardyab33a172005-08-14 19:31:36 -0700412 if (!sk)
413 return -ENOMEM;
414
415 sock_init_data(sock, sk);
416
417 nlk = nlk_sk(sk);
Patrick McHardyffa4d722007-04-25 14:01:17 -0700418 if (cb_mutex)
419 nlk->cb_mutex = cb_mutex;
420 else {
421 nlk->cb_mutex = &nlk->cb_def_mutex;
422 mutex_init(nlk->cb_mutex);
423 }
Patrick McHardyab33a172005-08-14 19:31:36 -0700424 init_waitqueue_head(&nlk->wait);
425
426 sk->sk_destruct = netlink_sock_destruct;
427 sk->sk_protocol = protocol;
428 return 0;
429}
430
Eric Paris3f378b62009-11-05 22:18:14 -0800431static int netlink_create(struct net *net, struct socket *sock, int protocol,
432 int kern)
Patrick McHardyab33a172005-08-14 19:31:36 -0700433{
434 struct module *module = NULL;
Patrick McHardyaf65bdf2007-04-20 14:14:21 -0700435 struct mutex *cb_mutex;
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700436 struct netlink_sock *nlk;
Patrick McHardyab33a172005-08-14 19:31:36 -0700437 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
439 sock->state = SS_UNCONNECTED;
440
441 if (sock->type != SOCK_RAW && sock->type != SOCK_DGRAM)
442 return -ESOCKTNOSUPPORT;
443
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800444 if (protocol < 0 || protocol >= MAX_LINKS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 return -EPROTONOSUPPORT;
446
Patrick McHardy77247bb2005-08-14 19:27:13 -0700447 netlink_lock_table();
Johannes Berg95a5afc2008-10-16 15:24:51 -0700448#ifdef CONFIG_MODULES
Patrick McHardyab33a172005-08-14 19:31:36 -0700449 if (!nl_table[protocol].registered) {
Patrick McHardy77247bb2005-08-14 19:27:13 -0700450 netlink_unlock_table();
Harald Welte4fdb3bb2005-08-09 19:40:55 -0700451 request_module("net-pf-%d-proto-%d", PF_NETLINK, protocol);
Patrick McHardy77247bb2005-08-14 19:27:13 -0700452 netlink_lock_table();
Harald Welte4fdb3bb2005-08-09 19:40:55 -0700453 }
Patrick McHardyab33a172005-08-14 19:31:36 -0700454#endif
455 if (nl_table[protocol].registered &&
456 try_module_get(nl_table[protocol].module))
457 module = nl_table[protocol].module;
Alexey Dobriyan974c37e2010-01-30 10:05:05 +0000458 else
459 err = -EPROTONOSUPPORT;
Patrick McHardyaf65bdf2007-04-20 14:14:21 -0700460 cb_mutex = nl_table[protocol].cb_mutex;
Patrick McHardy77247bb2005-08-14 19:27:13 -0700461 netlink_unlock_table();
Harald Welte4fdb3bb2005-08-09 19:40:55 -0700462
Alexey Dobriyan974c37e2010-01-30 10:05:05 +0000463 if (err < 0)
464 goto out;
465
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800466 err = __netlink_create(net, sock, cb_mutex, protocol);
467 if (err < 0)
Patrick McHardyab33a172005-08-14 19:31:36 -0700468 goto out_module;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469
David S. Miller6f756a82008-11-23 17:34:03 -0800470 local_bh_disable();
Eric Dumazetc1fd3b92008-11-23 15:48:22 -0800471 sock_prot_inuse_add(net, &netlink_proto, 1);
David S. Miller6f756a82008-11-23 17:34:03 -0800472 local_bh_enable();
473
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700474 nlk = nlk_sk(sock->sk);
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700475 nlk->module = module;
Patrick McHardyab33a172005-08-14 19:31:36 -0700476out:
477 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
Patrick McHardyab33a172005-08-14 19:31:36 -0700479out_module:
480 module_put(module);
481 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482}
483
484static int netlink_release(struct socket *sock)
485{
486 struct sock *sk = sock->sk;
487 struct netlink_sock *nlk;
488
489 if (!sk)
490 return 0;
491
492 netlink_remove(sk);
Denis Lunevac57b3a2007-04-18 17:05:58 -0700493 sock_orphan(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 nlk = nlk_sk(sk);
495
Herbert Xu3f660d62007-05-03 03:17:14 -0700496 /*
497 * OK. Socket is unlinked, any packets that arrive now
498 * will be purged.
499 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 sock->sk = NULL;
502 wake_up_interruptible_all(&nlk->wait);
503
504 skb_queue_purge(&sk->sk_write_queue);
505
Johannes Berg649300b2009-11-16 12:05:34 +0000506 if (nlk->pid) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 struct netlink_notify n = {
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900508 .net = sock_net(sk),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 .protocol = sk->sk_protocol,
510 .pid = nlk->pid,
511 };
Alan Sterne041c682006-03-27 01:16:30 -0800512 atomic_notifier_call_chain(&netlink_chain,
513 NETLINK_URELEASE, &n);
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +0900514 }
Harald Welte4fdb3bb2005-08-09 19:40:55 -0700515
Mariusz Kozlowski5e7c0012007-01-02 15:24:30 -0800516 module_put(nlk->module);
Harald Welte4fdb3bb2005-08-09 19:40:55 -0700517
Patrick McHardy4277a082006-03-20 18:52:01 -0800518 netlink_table_grab();
Denis V. Lunevaed81562007-10-10 21:14:32 -0700519 if (netlink_is_kernel(sk)) {
Denis V. Lunev869e58f2008-01-18 23:53:31 -0800520 BUG_ON(nl_table[sk->sk_protocol].registered == 0);
521 if (--nl_table[sk->sk_protocol].registered == 0) {
522 kfree(nl_table[sk->sk_protocol].listeners);
523 nl_table[sk->sk_protocol].module = NULL;
524 nl_table[sk->sk_protocol].registered = 0;
525 }
Patrick McHardy4277a082006-03-20 18:52:01 -0800526 } else if (nlk->subscriptions)
527 netlink_update_listeners(sk);
528 netlink_table_ungrab();
Patrick McHardy77247bb2005-08-14 19:27:13 -0700529
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700530 kfree(nlk->groups);
531 nlk->groups = NULL;
532
Eric Dumazet37558102008-11-24 14:05:22 -0800533 local_bh_disable();
Eric Dumazetc1fd3b92008-11-23 15:48:22 -0800534 sock_prot_inuse_add(sock_net(sk), &netlink_proto, -1);
Eric Dumazet37558102008-11-24 14:05:22 -0800535 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 sock_put(sk);
537 return 0;
538}
539
540static int netlink_autobind(struct socket *sock)
541{
542 struct sock *sk = sock->sk;
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900543 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 struct nl_pid_hash *hash = &nl_table[sk->sk_protocol].hash;
545 struct hlist_head *head;
546 struct sock *osk;
547 struct hlist_node *node;
Tom Goff66aa4a52010-03-19 15:38:50 +0000548 s32 pid = task_tgid_vnr(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 int err;
550 static s32 rover = -4097;
551
552retry:
553 cond_resched();
554 netlink_table_grab();
555 head = nl_pid_hashfn(hash, pid);
556 sk_for_each(osk, node, head) {
YOSHIFUJI Hideaki878628f2008-03-26 03:57:35 +0900557 if (!net_eq(sock_net(osk), net))
Eric W. Biedermanb4b51022007-09-12 13:05:38 +0200558 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 if (nlk_sk(osk)->pid == pid) {
560 /* Bind collision, search negative pid values. */
561 pid = rover--;
562 if (rover > -4097)
563 rover = -4097;
564 netlink_table_ungrab();
565 goto retry;
566 }
567 }
568 netlink_table_ungrab();
569
Eric W. Biedermanb4b51022007-09-12 13:05:38 +0200570 err = netlink_insert(sk, net, pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 if (err == -EADDRINUSE)
572 goto retry;
David S. Millerd470e3b2005-06-26 15:31:51 -0700573
574 /* If 2 threads race to autobind, that is fine. */
575 if (err == -EBUSY)
576 err = 0;
577
578 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579}
580
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +0900581static inline int netlink_capable(struct socket *sock, unsigned int flag)
582{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 return (nl_table[sock->sk->sk_protocol].nl_nonroot & flag) ||
584 capable(CAP_NET_ADMIN);
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +0900585}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700587static void
588netlink_update_subscriptions(struct sock *sk, unsigned int subscriptions)
589{
590 struct netlink_sock *nlk = nlk_sk(sk);
591
592 if (nlk->subscriptions && !subscriptions)
593 __sk_del_bind_node(sk);
594 else if (!nlk->subscriptions && subscriptions)
595 sk_add_bind_node(sk, &nl_table[sk->sk_protocol].mc_list);
596 nlk->subscriptions = subscriptions;
597}
598
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700599static int netlink_realloc_groups(struct sock *sk)
Patrick McHardy513c2502005-09-06 15:43:59 -0700600{
601 struct netlink_sock *nlk = nlk_sk(sk);
602 unsigned int groups;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700603 unsigned long *new_groups;
Patrick McHardy513c2502005-09-06 15:43:59 -0700604 int err = 0;
605
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700606 netlink_table_grab();
607
Patrick McHardy513c2502005-09-06 15:43:59 -0700608 groups = nl_table[sk->sk_protocol].groups;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700609 if (!nl_table[sk->sk_protocol].registered) {
Patrick McHardy513c2502005-09-06 15:43:59 -0700610 err = -ENOENT;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700611 goto out_unlock;
612 }
Patrick McHardy513c2502005-09-06 15:43:59 -0700613
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700614 if (nlk->ngroups >= groups)
615 goto out_unlock;
Patrick McHardy513c2502005-09-06 15:43:59 -0700616
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700617 new_groups = krealloc(nlk->groups, NLGRPSZ(groups), GFP_ATOMIC);
618 if (new_groups == NULL) {
619 err = -ENOMEM;
620 goto out_unlock;
621 }
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800622 memset((char *)new_groups + NLGRPSZ(nlk->ngroups), 0,
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700623 NLGRPSZ(groups) - NLGRPSZ(nlk->ngroups));
624
625 nlk->groups = new_groups;
Patrick McHardy513c2502005-09-06 15:43:59 -0700626 nlk->ngroups = groups;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700627 out_unlock:
628 netlink_table_ungrab();
629 return err;
Patrick McHardy513c2502005-09-06 15:43:59 -0700630}
631
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800632static int netlink_bind(struct socket *sock, struct sockaddr *addr,
633 int addr_len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634{
635 struct sock *sk = sock->sk;
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900636 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 struct netlink_sock *nlk = nlk_sk(sk);
638 struct sockaddr_nl *nladdr = (struct sockaddr_nl *)addr;
639 int err;
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +0900640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 if (nladdr->nl_family != AF_NETLINK)
642 return -EINVAL;
643
644 /* Only superuser is allowed to listen multicasts */
Patrick McHardy513c2502005-09-06 15:43:59 -0700645 if (nladdr->nl_groups) {
646 if (!netlink_capable(sock, NL_NONROOT_RECV))
647 return -EPERM;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700648 err = netlink_realloc_groups(sk);
649 if (err)
650 return err;
Patrick McHardy513c2502005-09-06 15:43:59 -0700651 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652
653 if (nlk->pid) {
654 if (nladdr->nl_pid != nlk->pid)
655 return -EINVAL;
656 } else {
657 err = nladdr->nl_pid ?
Eric W. Biedermanb4b51022007-09-12 13:05:38 +0200658 netlink_insert(sk, net, nladdr->nl_pid) :
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 netlink_autobind(sock);
660 if (err)
661 return err;
662 }
663
Patrick McHardy513c2502005-09-06 15:43:59 -0700664 if (!nladdr->nl_groups && (nlk->groups == NULL || !(u32)nlk->groups[0]))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 return 0;
666
667 netlink_table_grab();
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700668 netlink_update_subscriptions(sk, nlk->subscriptions +
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +0900669 hweight32(nladdr->nl_groups) -
670 hweight32(nlk->groups[0]));
671 nlk->groups[0] = (nlk->groups[0] & ~0xffffffffUL) | nladdr->nl_groups;
Patrick McHardy4277a082006-03-20 18:52:01 -0800672 netlink_update_listeners(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 netlink_table_ungrab();
674
675 return 0;
676}
677
678static int netlink_connect(struct socket *sock, struct sockaddr *addr,
679 int alen, int flags)
680{
681 int err = 0;
682 struct sock *sk = sock->sk;
683 struct netlink_sock *nlk = nlk_sk(sk);
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800684 struct sockaddr_nl *nladdr = (struct sockaddr_nl *)addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685
Changli Gao6503d962010-03-31 22:58:26 +0000686 if (alen < sizeof(addr->sa_family))
687 return -EINVAL;
688
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 if (addr->sa_family == AF_UNSPEC) {
690 sk->sk_state = NETLINK_UNCONNECTED;
691 nlk->dst_pid = 0;
Patrick McHardyd629b832005-08-14 19:27:50 -0700692 nlk->dst_group = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 return 0;
694 }
695 if (addr->sa_family != AF_NETLINK)
696 return -EINVAL;
697
698 /* Only superuser is allowed to send multicasts */
699 if (nladdr->nl_groups && !netlink_capable(sock, NL_NONROOT_SEND))
700 return -EPERM;
701
702 if (!nlk->pid)
703 err = netlink_autobind(sock);
704
705 if (err == 0) {
706 sk->sk_state = NETLINK_CONNECTED;
707 nlk->dst_pid = nladdr->nl_pid;
Patrick McHardyd629b832005-08-14 19:27:50 -0700708 nlk->dst_group = ffs(nladdr->nl_groups);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 }
710
711 return err;
712}
713
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800714static int netlink_getname(struct socket *sock, struct sockaddr *addr,
715 int *addr_len, int peer)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716{
717 struct sock *sk = sock->sk;
718 struct netlink_sock *nlk = nlk_sk(sk);
Cyrill Gorcunov13cfa972009-11-08 05:51:19 +0000719 DECLARE_SOCKADDR(struct sockaddr_nl *, nladdr, addr);
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +0900720
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 nladdr->nl_family = AF_NETLINK;
722 nladdr->nl_pad = 0;
723 *addr_len = sizeof(*nladdr);
724
725 if (peer) {
726 nladdr->nl_pid = nlk->dst_pid;
Patrick McHardyd629b832005-08-14 19:27:50 -0700727 nladdr->nl_groups = netlink_group_mask(nlk->dst_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 } else {
729 nladdr->nl_pid = nlk->pid;
Patrick McHardy513c2502005-09-06 15:43:59 -0700730 nladdr->nl_groups = nlk->groups ? nlk->groups[0] : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 }
732 return 0;
733}
734
735static void netlink_overrun(struct sock *sk)
736{
Pablo Neira Ayuso38938bf2009-03-24 16:37:55 -0700737 struct netlink_sock *nlk = nlk_sk(sk);
738
739 if (!(nlk->flags & NETLINK_RECV_NO_ENOBUFS)) {
740 if (!test_and_set_bit(0, &nlk_sk(sk)->state)) {
741 sk->sk_err = ENOBUFS;
742 sk->sk_error_report(sk);
743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 }
Pablo Neira Ayuso38938bf2009-03-24 16:37:55 -0700745 atomic_inc(&sk->sk_drops);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746}
747
748static struct sock *netlink_getsockbypid(struct sock *ssk, u32 pid)
749{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 struct sock *sock;
751 struct netlink_sock *nlk;
752
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +0900753 sock = netlink_lookup(sock_net(ssk), ssk->sk_protocol, pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 if (!sock)
755 return ERR_PTR(-ECONNREFUSED);
756
757 /* Don't bother queuing skb if kernel socket has no input function */
758 nlk = nlk_sk(sock);
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -0700759 if (sock->sk_state == NETLINK_CONNECTED &&
760 nlk->dst_pid != nlk_sk(ssk)->pid) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 sock_put(sock);
762 return ERR_PTR(-ECONNREFUSED);
763 }
764 return sock;
765}
766
767struct sock *netlink_getsockbyfilp(struct file *filp)
768{
Josef Sipek6db5fc52006-12-08 02:37:25 -0800769 struct inode *inode = filp->f_path.dentry->d_inode;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 struct sock *sock;
771
772 if (!S_ISSOCK(inode->i_mode))
773 return ERR_PTR(-ENOTSOCK);
774
775 sock = SOCKET_I(inode)->sk;
776 if (sock->sk_family != AF_NETLINK)
777 return ERR_PTR(-EINVAL);
778
779 sock_hold(sock);
780 return sock;
781}
782
783/*
784 * Attach a skb to a netlink socket.
785 * The caller must hold a reference to the destination socket. On error, the
786 * reference is dropped. The skb is not send to the destination, just all
787 * all error checks are performed and memory in the queue is reserved.
788 * Return values:
789 * < 0: error. skb freed, reference to sock dropped.
790 * 0: continue
791 * 1: repeat lookup - reference dropped while waiting for socket memory.
792 */
Denis V. Lunev9457afe2008-06-05 11:23:39 -0700793int netlink_attachskb(struct sock *sk, struct sk_buff *skb,
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -0800794 long *timeo, struct sock *ssk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795{
796 struct netlink_sock *nlk;
797
798 nlk = nlk_sk(sk);
799
800 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
801 test_bit(0, &nlk->state)) {
802 DECLARE_WAITQUEUE(wait, current);
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -0800803 if (!*timeo) {
Denis V. Lunevaed81562007-10-10 21:14:32 -0700804 if (!ssk || netlink_is_kernel(ssk))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 netlink_overrun(sk);
806 sock_put(sk);
807 kfree_skb(skb);
808 return -EAGAIN;
809 }
810
811 __set_current_state(TASK_INTERRUPTIBLE);
812 add_wait_queue(&nlk->wait, &wait);
813
814 if ((atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
815 test_bit(0, &nlk->state)) &&
816 !sock_flag(sk, SOCK_DEAD))
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -0800817 *timeo = schedule_timeout(*timeo);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818
819 __set_current_state(TASK_RUNNING);
820 remove_wait_queue(&nlk->wait, &wait);
821 sock_put(sk);
822
823 if (signal_pending(current)) {
824 kfree_skb(skb);
Patrick McHardyc3d8d1e2007-11-07 02:42:09 -0800825 return sock_intr_errno(*timeo);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 }
827 return 1;
828 }
829 skb_set_owner_r(skb, sk);
830 return 0;
831}
832
Denis V. Lunev7ee015e2007-10-10 21:14:03 -0700833int netlink_sendskb(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 int len = skb->len;
836
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 skb_queue_tail(&sk->sk_receive_queue, skb);
838 sk->sk_data_ready(sk, len);
839 sock_put(sk);
840 return len;
841}
842
843void netlink_detachskb(struct sock *sk, struct sk_buff *skb)
844{
845 kfree_skb(skb);
846 sock_put(sk);
847}
848
Victor Fusco37da6472005-07-18 13:35:43 -0700849static inline struct sk_buff *netlink_trim(struct sk_buff *skb,
Al Virodd0fc662005-10-07 07:46:04 +0100850 gfp_t allocation)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851{
852 int delta;
853
854 skb_orphan(skb);
855
Arnaldo Carvalho de Melo4305b542007-04-19 20:43:29 -0700856 delta = skb->end - skb->tail;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 if (delta * 2 < skb->truesize)
858 return skb;
859
860 if (skb_shared(skb)) {
861 struct sk_buff *nskb = skb_clone(skb, allocation);
862 if (!nskb)
863 return skb;
864 kfree_skb(skb);
865 skb = nskb;
866 }
867
868 if (!pskb_expand_head(skb, 0, -delta, allocation))
869 skb->truesize -= delta;
870
871 return skb;
872}
873
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -0700874static inline void netlink_rcv_wake(struct sock *sk)
875{
876 struct netlink_sock *nlk = nlk_sk(sk);
877
878 if (skb_queue_empty(&sk->sk_receive_queue))
879 clear_bit(0, &nlk->state);
880 if (!test_bit(0, &nlk->state))
881 wake_up_interruptible(&nlk->wait);
882}
883
884static inline int netlink_unicast_kernel(struct sock *sk, struct sk_buff *skb)
885{
886 int ret;
887 struct netlink_sock *nlk = nlk_sk(sk);
888
889 ret = -ECONNREFUSED;
890 if (nlk->netlink_rcv != NULL) {
891 ret = skb->len;
892 skb_set_owner_r(skb, sk);
893 nlk->netlink_rcv(skb);
894 }
895 kfree_skb(skb);
896 sock_put(sk);
897 return ret;
898}
899
900int netlink_unicast(struct sock *ssk, struct sk_buff *skb,
901 u32 pid, int nonblock)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902{
903 struct sock *sk;
904 int err;
905 long timeo;
906
907 skb = netlink_trim(skb, gfp_any());
908
909 timeo = sock_sndtimeo(ssk, nonblock);
910retry:
911 sk = netlink_getsockbypid(ssk, pid);
912 if (IS_ERR(sk)) {
913 kfree_skb(skb);
914 return PTR_ERR(sk);
915 }
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -0700916 if (netlink_is_kernel(sk))
917 return netlink_unicast_kernel(sk, skb);
918
Stephen Hemmingerb1153f22008-03-21 15:46:12 -0700919 if (sk_filter(sk, skb)) {
Wang Chen84874602008-07-01 19:55:09 -0700920 err = skb->len;
Stephen Hemmingerb1153f22008-03-21 15:46:12 -0700921 kfree_skb(skb);
922 sock_put(sk);
923 return err;
924 }
925
Denis V. Lunev9457afe2008-06-05 11:23:39 -0700926 err = netlink_attachskb(sk, skb, &timeo, ssk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 if (err == 1)
928 goto retry;
929 if (err)
930 return err;
931
Denis V. Lunev7ee015e2007-10-10 21:14:03 -0700932 return netlink_sendskb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933}
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800934EXPORT_SYMBOL(netlink_unicast);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
Patrick McHardy4277a082006-03-20 18:52:01 -0800936int netlink_has_listeners(struct sock *sk, unsigned int group)
937{
938 int res = 0;
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700939 unsigned long *listeners;
Patrick McHardy4277a082006-03-20 18:52:01 -0800940
Denis V. Lunevaed81562007-10-10 21:14:32 -0700941 BUG_ON(!netlink_is_kernel(sk));
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700942
943 rcu_read_lock();
944 listeners = rcu_dereference(nl_table[sk->sk_protocol].listeners);
945
Patrick McHardy4277a082006-03-20 18:52:01 -0800946 if (group - 1 < nl_table[sk->sk_protocol].groups)
Johannes Bergb4ff4f02007-07-18 15:46:06 -0700947 res = test_bit(group - 1, listeners);
948
949 rcu_read_unlock();
950
Patrick McHardy4277a082006-03-20 18:52:01 -0800951 return res;
952}
953EXPORT_SYMBOL_GPL(netlink_has_listeners);
954
Patrick McHardy6ac552f2007-12-04 00:19:38 -0800955static inline int netlink_broadcast_deliver(struct sock *sk,
956 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957{
958 struct netlink_sock *nlk = nlk_sk(sk);
959
960 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
961 !test_bit(0, &nlk->state)) {
962 skb_set_owner_r(skb, sk);
963 skb_queue_tail(&sk->sk_receive_queue, skb);
964 sk->sk_data_ready(sk, skb->len);
965 return atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf;
966 }
967 return -1;
968}
969
970struct netlink_broadcast_data {
971 struct sock *exclude_sk;
Eric W. Biedermanb4b51022007-09-12 13:05:38 +0200972 struct net *net;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 u32 pid;
974 u32 group;
975 int failure;
Pablo Neira Ayusoff491a72009-02-05 23:56:36 -0800976 int delivery_failure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 int congested;
978 int delivered;
Al Viro7d877f32005-10-21 03:20:43 -0400979 gfp_t allocation;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 struct sk_buff *skb, *skb2;
Eric W. Biederman910a7e92010-05-04 17:36:46 -0700981 int (*tx_filter)(struct sock *dsk, struct sk_buff *skb, void *data);
982 void *tx_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983};
984
985static inline int do_one_broadcast(struct sock *sk,
986 struct netlink_broadcast_data *p)
987{
988 struct netlink_sock *nlk = nlk_sk(sk);
989 int val;
990
991 if (p->exclude_sk == sk)
992 goto out;
993
Patrick McHardyf7fa9b12005-08-15 12:29:13 -0700994 if (nlk->pid == p->pid || p->group - 1 >= nlk->ngroups ||
995 !test_bit(p->group - 1, nlk->groups))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 goto out;
997
YOSHIFUJI Hideaki878628f2008-03-26 03:57:35 +0900998 if (!net_eq(sock_net(sk), p->net))
Eric W. Biedermanb4b51022007-09-12 13:05:38 +0200999 goto out;
1000
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 if (p->failure) {
1002 netlink_overrun(sk);
1003 goto out;
1004 }
1005
1006 sock_hold(sk);
1007 if (p->skb2 == NULL) {
Tommy S. Christensen68acc022005-05-19 13:06:35 -07001008 if (skb_shared(p->skb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 p->skb2 = skb_clone(p->skb, p->allocation);
1010 } else {
Tommy S. Christensen68acc022005-05-19 13:06:35 -07001011 p->skb2 = skb_get(p->skb);
1012 /*
1013 * skb ownership may have been set when
1014 * delivered to a previous socket.
1015 */
1016 skb_orphan(p->skb2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 }
1018 }
1019 if (p->skb2 == NULL) {
1020 netlink_overrun(sk);
1021 /* Clone failed. Notify ALL listeners. */
1022 p->failure = 1;
Pablo Neira Ayusobe0c22a2009-02-18 01:40:43 +00001023 if (nlk->flags & NETLINK_BROADCAST_SEND_ERROR)
1024 p->delivery_failure = 1;
Eric W. Biederman910a7e92010-05-04 17:36:46 -07001025 } else if (p->tx_filter && p->tx_filter(sk, p->skb2, p->tx_data)) {
1026 kfree_skb(p->skb2);
1027 p->skb2 = NULL;
Stephen Hemmingerb1153f22008-03-21 15:46:12 -07001028 } else if (sk_filter(sk, p->skb2)) {
1029 kfree_skb(p->skb2);
1030 p->skb2 = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 } else if ((val = netlink_broadcast_deliver(sk, p->skb2)) < 0) {
1032 netlink_overrun(sk);
Pablo Neira Ayusobe0c22a2009-02-18 01:40:43 +00001033 if (nlk->flags & NETLINK_BROADCAST_SEND_ERROR)
1034 p->delivery_failure = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 } else {
1036 p->congested |= val;
1037 p->delivered = 1;
1038 p->skb2 = NULL;
1039 }
1040 sock_put(sk);
1041
1042out:
1043 return 0;
1044}
1045
Eric W. Biederman910a7e92010-05-04 17:36:46 -07001046int netlink_broadcast_filtered(struct sock *ssk, struct sk_buff *skb, u32 pid,
1047 u32 group, gfp_t allocation,
1048 int (*filter)(struct sock *dsk, struct sk_buff *skb, void *data),
1049 void *filter_data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050{
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +09001051 struct net *net = sock_net(ssk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 struct netlink_broadcast_data info;
1053 struct hlist_node *node;
1054 struct sock *sk;
1055
1056 skb = netlink_trim(skb, allocation);
1057
1058 info.exclude_sk = ssk;
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001059 info.net = net;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060 info.pid = pid;
1061 info.group = group;
1062 info.failure = 0;
Pablo Neira Ayusoff491a72009-02-05 23:56:36 -08001063 info.delivery_failure = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 info.congested = 0;
1065 info.delivered = 0;
1066 info.allocation = allocation;
1067 info.skb = skb;
1068 info.skb2 = NULL;
Eric W. Biederman910a7e92010-05-04 17:36:46 -07001069 info.tx_filter = filter;
1070 info.tx_data = filter_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
1072 /* While we sleep in clone, do not allow to change socket list */
1073
1074 netlink_lock_table();
1075
1076 sk_for_each_bound(sk, node, &nl_table[ssk->sk_protocol].mc_list)
1077 do_one_broadcast(sk, &info);
1078
Neil Horman70d4bf62010-07-20 06:45:56 +00001079 consume_skb(skb);
Tommy S. Christensenaa1c6a62005-05-19 13:07:32 -07001080
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 netlink_unlock_table();
1082
Neil Horman70d4bf62010-07-20 06:45:56 +00001083 if (info.delivery_failure) {
1084 kfree_skb(info.skb2);
Pablo Neira Ayusoff491a72009-02-05 23:56:36 -08001085 return -ENOBUFS;
Neil Horman70d4bf62010-07-20 06:45:56 +00001086 } else
1087 consume_skb(info.skb2);
Pablo Neira Ayusoff491a72009-02-05 23:56:36 -08001088
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 if (info.delivered) {
1090 if (info.congested && (allocation & __GFP_WAIT))
1091 yield();
1092 return 0;
1093 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 return -ESRCH;
1095}
Eric W. Biederman910a7e92010-05-04 17:36:46 -07001096EXPORT_SYMBOL(netlink_broadcast_filtered);
1097
1098int netlink_broadcast(struct sock *ssk, struct sk_buff *skb, u32 pid,
1099 u32 group, gfp_t allocation)
1100{
1101 return netlink_broadcast_filtered(ssk, skb, pid, group, allocation,
1102 NULL, NULL);
1103}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001104EXPORT_SYMBOL(netlink_broadcast);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105
1106struct netlink_set_err_data {
1107 struct sock *exclude_sk;
1108 u32 pid;
1109 u32 group;
1110 int code;
1111};
1112
1113static inline int do_one_set_err(struct sock *sk,
1114 struct netlink_set_err_data *p)
1115{
1116 struct netlink_sock *nlk = nlk_sk(sk);
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001117 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
1119 if (sk == p->exclude_sk)
1120 goto out;
1121
Octavian Purdila09ad9bc2009-11-25 15:14:13 -08001122 if (!net_eq(sock_net(sk), sock_net(p->exclude_sk)))
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001123 goto out;
1124
Patrick McHardyf7fa9b12005-08-15 12:29:13 -07001125 if (nlk->pid == p->pid || p->group - 1 >= nlk->ngroups ||
1126 !test_bit(p->group - 1, nlk->groups))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 goto out;
1128
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001129 if (p->code == ENOBUFS && nlk->flags & NETLINK_RECV_NO_ENOBUFS) {
1130 ret = 1;
1131 goto out;
1132 }
1133
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 sk->sk_err = p->code;
1135 sk->sk_error_report(sk);
1136out:
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001137 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138}
1139
Pablo Neira Ayuso4843b932009-03-03 23:37:30 -08001140/**
1141 * netlink_set_err - report error to broadcast listeners
1142 * @ssk: the kernel netlink socket, as returned by netlink_kernel_create()
1143 * @pid: the PID of a process that we want to skip (if any)
1144 * @groups: the broadcast group that will notice the error
1145 * @code: error code, must be negative (as usual in kernelspace)
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001146 *
1147 * This function returns the number of broadcast listeners that have set the
1148 * NETLINK_RECV_NO_ENOBUFS socket option.
Pablo Neira Ayuso4843b932009-03-03 23:37:30 -08001149 */
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001150int netlink_set_err(struct sock *ssk, u32 pid, u32 group, int code)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151{
1152 struct netlink_set_err_data info;
1153 struct hlist_node *node;
1154 struct sock *sk;
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001155 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156
1157 info.exclude_sk = ssk;
1158 info.pid = pid;
1159 info.group = group;
Pablo Neira Ayuso4843b932009-03-03 23:37:30 -08001160 /* sk->sk_err wants a positive error value */
1161 info.code = -code;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162
1163 read_lock(&nl_table_lock);
1164
1165 sk_for_each_bound(sk, node, &nl_table[ssk->sk_protocol].mc_list)
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001166 ret += do_one_set_err(sk, &info);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
1168 read_unlock(&nl_table_lock);
Pablo Neira Ayuso1a503072010-03-18 14:24:42 +00001169 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170}
Pablo Neira Ayusodd5b6ce2009-03-23 13:21:06 +01001171EXPORT_SYMBOL(netlink_set_err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
Johannes Berg84659eb2007-07-18 15:47:05 -07001173/* must be called with netlink table grabbed */
1174static void netlink_update_socket_mc(struct netlink_sock *nlk,
1175 unsigned int group,
1176 int is_new)
1177{
1178 int old, new = !!is_new, subscriptions;
1179
1180 old = test_bit(group - 1, nlk->groups);
1181 subscriptions = nlk->subscriptions - old + new;
1182 if (new)
1183 __set_bit(group - 1, nlk->groups);
1184 else
1185 __clear_bit(group - 1, nlk->groups);
1186 netlink_update_subscriptions(&nlk->sk, subscriptions);
1187 netlink_update_listeners(&nlk->sk);
1188}
1189
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001190static int netlink_setsockopt(struct socket *sock, int level, int optname,
David S. Millerb7058842009-09-30 16:12:20 -07001191 char __user *optval, unsigned int optlen)
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001192{
1193 struct sock *sk = sock->sk;
1194 struct netlink_sock *nlk = nlk_sk(sk);
Johannes Bergeb496532007-07-18 02:07:51 -07001195 unsigned int val = 0;
1196 int err;
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001197
1198 if (level != SOL_NETLINK)
1199 return -ENOPROTOOPT;
1200
1201 if (optlen >= sizeof(int) &&
Johannes Bergeb496532007-07-18 02:07:51 -07001202 get_user(val, (unsigned int __user *)optval))
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001203 return -EFAULT;
1204
1205 switch (optname) {
1206 case NETLINK_PKTINFO:
1207 if (val)
1208 nlk->flags |= NETLINK_RECV_PKTINFO;
1209 else
1210 nlk->flags &= ~NETLINK_RECV_PKTINFO;
1211 err = 0;
1212 break;
1213 case NETLINK_ADD_MEMBERSHIP:
1214 case NETLINK_DROP_MEMBERSHIP: {
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001215 if (!netlink_capable(sock, NL_NONROOT_RECV))
1216 return -EPERM;
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001217 err = netlink_realloc_groups(sk);
1218 if (err)
1219 return err;
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001220 if (!val || val - 1 >= nlk->ngroups)
1221 return -EINVAL;
1222 netlink_table_grab();
Johannes Berg84659eb2007-07-18 15:47:05 -07001223 netlink_update_socket_mc(nlk, val,
1224 optname == NETLINK_ADD_MEMBERSHIP);
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001225 netlink_table_ungrab();
1226 err = 0;
1227 break;
1228 }
Pablo Neira Ayusobe0c22a2009-02-18 01:40:43 +00001229 case NETLINK_BROADCAST_ERROR:
1230 if (val)
1231 nlk->flags |= NETLINK_BROADCAST_SEND_ERROR;
1232 else
1233 nlk->flags &= ~NETLINK_BROADCAST_SEND_ERROR;
1234 err = 0;
1235 break;
Pablo Neira Ayuso38938bf2009-03-24 16:37:55 -07001236 case NETLINK_NO_ENOBUFS:
1237 if (val) {
1238 nlk->flags |= NETLINK_RECV_NO_ENOBUFS;
1239 clear_bit(0, &nlk->state);
1240 wake_up_interruptible(&nlk->wait);
1241 } else
1242 nlk->flags &= ~NETLINK_RECV_NO_ENOBUFS;
1243 err = 0;
1244 break;
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001245 default:
1246 err = -ENOPROTOOPT;
1247 }
1248 return err;
1249}
1250
1251static int netlink_getsockopt(struct socket *sock, int level, int optname,
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09001252 char __user *optval, int __user *optlen)
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001253{
1254 struct sock *sk = sock->sk;
1255 struct netlink_sock *nlk = nlk_sk(sk);
1256 int len, val, err;
1257
1258 if (level != SOL_NETLINK)
1259 return -ENOPROTOOPT;
1260
1261 if (get_user(len, optlen))
1262 return -EFAULT;
1263 if (len < 0)
1264 return -EINVAL;
1265
1266 switch (optname) {
1267 case NETLINK_PKTINFO:
1268 if (len < sizeof(int))
1269 return -EINVAL;
1270 len = sizeof(int);
1271 val = nlk->flags & NETLINK_RECV_PKTINFO ? 1 : 0;
Heiko Carstensa27b58f2006-10-30 15:06:12 -08001272 if (put_user(len, optlen) ||
1273 put_user(val, optval))
1274 return -EFAULT;
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001275 err = 0;
1276 break;
Pablo Neira Ayusobe0c22a2009-02-18 01:40:43 +00001277 case NETLINK_BROADCAST_ERROR:
1278 if (len < sizeof(int))
1279 return -EINVAL;
1280 len = sizeof(int);
1281 val = nlk->flags & NETLINK_BROADCAST_SEND_ERROR ? 1 : 0;
1282 if (put_user(len, optlen) ||
1283 put_user(val, optval))
1284 return -EFAULT;
1285 err = 0;
1286 break;
Pablo Neira Ayuso38938bf2009-03-24 16:37:55 -07001287 case NETLINK_NO_ENOBUFS:
1288 if (len < sizeof(int))
1289 return -EINVAL;
1290 len = sizeof(int);
1291 val = nlk->flags & NETLINK_RECV_NO_ENOBUFS ? 1 : 0;
1292 if (put_user(len, optlen) ||
1293 put_user(val, optval))
1294 return -EFAULT;
1295 err = 0;
1296 break;
Patrick McHardy9a4595b2005-08-15 12:32:15 -07001297 default:
1298 err = -ENOPROTOOPT;
1299 }
1300 return err;
1301}
1302
1303static void netlink_cmsg_recv_pktinfo(struct msghdr *msg, struct sk_buff *skb)
1304{
1305 struct nl_pktinfo info;
1306
1307 info.group = NETLINK_CB(skb).dst_group;
1308 put_cmsg(msg, SOL_NETLINK, NETLINK_PKTINFO, sizeof(info), &info);
1309}
1310
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311static int netlink_sendmsg(struct kiocb *kiocb, struct socket *sock,
1312 struct msghdr *msg, size_t len)
1313{
1314 struct sock_iocb *siocb = kiocb_to_siocb(kiocb);
1315 struct sock *sk = sock->sk;
1316 struct netlink_sock *nlk = nlk_sk(sk);
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001317 struct sockaddr_nl *addr = msg->msg_name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 u32 dst_pid;
Patrick McHardyd629b832005-08-14 19:27:50 -07001319 u32 dst_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 struct sk_buff *skb;
1321 int err;
1322 struct scm_cookie scm;
1323
1324 if (msg->msg_flags&MSG_OOB)
1325 return -EOPNOTSUPP;
1326
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001327 if (NULL == siocb->scm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 siocb->scm = &scm;
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001329 memset(&scm, 0, sizeof(scm));
1330 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 err = scm_send(sock, msg, siocb->scm);
1332 if (err < 0)
1333 return err;
1334
1335 if (msg->msg_namelen) {
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001336 err = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 if (addr->nl_family != AF_NETLINK)
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001338 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 dst_pid = addr->nl_pid;
Patrick McHardyd629b832005-08-14 19:27:50 -07001340 dst_group = ffs(addr->nl_groups);
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001341 err = -EPERM;
Patrick McHardyd629b832005-08-14 19:27:50 -07001342 if (dst_group && !netlink_capable(sock, NL_NONROOT_SEND))
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001343 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 } else {
1345 dst_pid = nlk->dst_pid;
Patrick McHardyd629b832005-08-14 19:27:50 -07001346 dst_group = nlk->dst_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 }
1348
1349 if (!nlk->pid) {
1350 err = netlink_autobind(sock);
1351 if (err)
1352 goto out;
1353 }
1354
1355 err = -EMSGSIZE;
1356 if (len > sk->sk_sndbuf - 32)
1357 goto out;
1358 err = -ENOBUFS;
Thomas Graf339bf982006-11-10 14:10:15 -08001359 skb = alloc_skb(len, GFP_KERNEL);
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001360 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 goto out;
1362
1363 NETLINK_CB(skb).pid = nlk->pid;
Patrick McHardyd629b832005-08-14 19:27:50 -07001364 NETLINK_CB(skb).dst_group = dst_group;
Al Viro0c11b942008-01-10 04:20:52 -05001365 NETLINK_CB(skb).loginuid = audit_get_loginuid(current);
Eric Paris25323862008-04-18 10:09:25 -04001366 NETLINK_CB(skb).sessionid = audit_get_sessionid(current);
Ahmed S. Darwish0ce784c2008-03-01 21:56:22 +02001367 security_task_getsecid(current, &(NETLINK_CB(skb).sid));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 memcpy(NETLINK_CREDS(skb), &siocb->scm->creds, sizeof(struct ucred));
1369
1370 /* What can I do? Netlink is asynchronous, so that
1371 we will have to save current capabilities to
1372 check them, when this message will be delivered
1373 to corresponding kernel module. --ANK (980802)
1374 */
1375
1376 err = -EFAULT;
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001377 if (memcpy_fromiovec(skb_put(skb, len), msg->msg_iov, len)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 kfree_skb(skb);
1379 goto out;
1380 }
1381
1382 err = security_netlink_send(sk, skb);
1383 if (err) {
1384 kfree_skb(skb);
1385 goto out;
1386 }
1387
Patrick McHardyd629b832005-08-14 19:27:50 -07001388 if (dst_group) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 atomic_inc(&skb->users);
Patrick McHardyd629b832005-08-14 19:27:50 -07001390 netlink_broadcast(sk, skb, dst_pid, dst_group, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 }
1392 err = netlink_unicast(sk, skb, dst_pid, msg->msg_flags&MSG_DONTWAIT);
1393
1394out:
Eric W. Biedermanb47030c2010-06-13 03:31:06 +00001395 scm_destroy(siocb->scm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 return err;
1397}
1398
1399static int netlink_recvmsg(struct kiocb *kiocb, struct socket *sock,
1400 struct msghdr *msg, size_t len,
1401 int flags)
1402{
1403 struct sock_iocb *siocb = kiocb_to_siocb(kiocb);
1404 struct scm_cookie scm;
1405 struct sock *sk = sock->sk;
1406 struct netlink_sock *nlk = nlk_sk(sk);
1407 int noblock = flags&MSG_DONTWAIT;
1408 size_t copied;
Eric Dumazet1235f502010-07-26 13:09:16 -07001409 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 int err;
1411
1412 if (flags&MSG_OOB)
1413 return -EOPNOTSUPP;
1414
1415 copied = 0;
1416
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001417 skb = skb_recv_datagram(sk, flags, noblock, &err);
1418 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419 goto out;
1420
Johannes Berg1dacc762009-07-01 11:26:02 +00001421#ifdef CONFIG_COMPAT_NETLINK_MESSAGES
1422 if (unlikely(skb_shinfo(skb)->frag_list)) {
1423 bool need_compat = !!(flags & MSG_CMSG_COMPAT);
1424
1425 /*
1426 * If this skb has a frag_list, then here that means that
1427 * we will have to use the frag_list skb for compat tasks
1428 * and the regular skb for non-compat tasks.
1429 *
1430 * The skb might (and likely will) be cloned, so we can't
1431 * just reset frag_list and go on with things -- we need to
1432 * keep that. For the compat case that's easy -- simply get
1433 * a reference to the compat skb and free the regular one
1434 * including the frag. For the non-compat case, we need to
1435 * avoid sending the frag to the user -- so assign NULL but
1436 * restore it below before freeing the skb.
1437 */
1438 if (need_compat) {
1439 struct sk_buff *compskb = skb_shinfo(skb)->frag_list;
1440 skb_get(compskb);
1441 kfree_skb(skb);
1442 skb = compskb;
1443 } else {
Eric Dumazet1235f502010-07-26 13:09:16 -07001444 /*
1445 * Before setting frag_list to NULL, we must get a
1446 * private copy of skb if shared (because of MSG_PEEK)
1447 */
1448 if (skb_shared(skb)) {
1449 struct sk_buff *nskb;
1450
1451 nskb = pskb_copy(skb, GFP_KERNEL);
1452 kfree_skb(skb);
1453 skb = nskb;
1454 err = -ENOMEM;
1455 if (!skb)
1456 goto out;
1457 }
1458 kfree_skb(skb_shinfo(skb)->frag_list);
Johannes Berg1dacc762009-07-01 11:26:02 +00001459 skb_shinfo(skb)->frag_list = NULL;
1460 }
1461 }
1462#endif
1463
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 msg->msg_namelen = 0;
1465
1466 copied = skb->len;
1467 if (len < copied) {
1468 msg->msg_flags |= MSG_TRUNC;
1469 copied = len;
1470 }
1471
Arnaldo Carvalho de Melobadff6d2007-03-13 13:06:52 -03001472 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, copied);
1474
1475 if (msg->msg_name) {
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001476 struct sockaddr_nl *addr = (struct sockaddr_nl *)msg->msg_name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 addr->nl_family = AF_NETLINK;
1478 addr->nl_pad = 0;
1479 addr->nl_pid = NETLINK_CB(skb).pid;
Patrick McHardyd629b832005-08-14 19:27:50 -07001480 addr->nl_groups = netlink_group_mask(NETLINK_CB(skb).dst_group);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 msg->msg_namelen = sizeof(*addr);
1482 }
1483
Patrick McHardycc9a06c2006-03-12 20:34:27 -08001484 if (nlk->flags & NETLINK_RECV_PKTINFO)
1485 netlink_cmsg_recv_pktinfo(msg, skb);
1486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 if (NULL == siocb->scm) {
1488 memset(&scm, 0, sizeof(scm));
1489 siocb->scm = &scm;
1490 }
1491 siocb->scm->creds = *NETLINK_CREDS(skb);
Patrick McHardy188ccb52007-05-03 03:27:01 -07001492 if (flags & MSG_TRUNC)
1493 copied = skb->len;
Johannes Berg1dacc762009-07-01 11:26:02 +00001494
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 skb_free_datagram(sk, skb);
1496
1497 if (nlk->cb && atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf / 2)
1498 netlink_dump(sk);
1499
1500 scm_recv(sock, msg, siocb->scm, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501out:
1502 netlink_rcv_wake(sk);
1503 return err ? : copied;
1504}
1505
1506static void netlink_data_ready(struct sock *sk, int len)
1507{
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -07001508 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509}
1510
1511/*
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09001512 * We export these functions to other modules. They provide a
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 * complete set of kernel non-blocking support for message
1514 * queueing.
1515 */
1516
1517struct sock *
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001518netlink_kernel_create(struct net *net, int unit, unsigned int groups,
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -07001519 void (*input)(struct sk_buff *skb),
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001520 struct mutex *cb_mutex, struct module *module)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521{
1522 struct socket *sock;
1523 struct sock *sk;
Patrick McHardy77247bb2005-08-14 19:27:13 -07001524 struct netlink_sock *nlk;
Patrick McHardy4277a082006-03-20 18:52:01 -08001525 unsigned long *listeners = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
Akinobu Mitafab2caf2006-08-29 02:15:24 -07001527 BUG_ON(!nl_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001529 if (unit < 0 || unit >= MAX_LINKS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 return NULL;
1531
1532 if (sock_create_lite(PF_NETLINK, SOCK_DGRAM, unit, &sock))
1533 return NULL;
1534
Pavel Emelyanov23fe1862008-01-30 19:31:06 -08001535 /*
1536 * We have to just have a reference on the net from sk, but don't
1537 * get_net it. Besides, we cannot get and then put the net here.
1538 * So we create one inside init_net and the move it to net.
1539 */
1540
1541 if (__netlink_create(&init_net, sock, cb_mutex, unit) < 0)
1542 goto out_sock_release_nosk;
1543
1544 sk = sock->sk;
Denis V. Lunevedf02082008-02-29 11:18:32 -08001545 sk_change_net(sk, net);
Harald Welte4fdb3bb2005-08-09 19:40:55 -07001546
Patrick McHardy4277a082006-03-20 18:52:01 -08001547 if (groups < 32)
1548 groups = 32;
1549
Johannes Berg6c04bb12009-07-10 09:51:32 +00001550 listeners = kzalloc(NLGRPSZ(groups) + sizeof(struct listeners_rcu_head),
1551 GFP_KERNEL);
Patrick McHardy4277a082006-03-20 18:52:01 -08001552 if (!listeners)
1553 goto out_sock_release;
1554
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 sk->sk_data_ready = netlink_data_ready;
1556 if (input)
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -07001557 nlk_sk(sk)->netlink_rcv = input;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001559 if (netlink_insert(sk, net, 0))
Patrick McHardy77247bb2005-08-14 19:27:13 -07001560 goto out_sock_release;
1561
1562 nlk = nlk_sk(sk);
1563 nlk->flags |= NETLINK_KERNEL_SOCKET;
1564
1565 netlink_table_grab();
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001566 if (!nl_table[unit].registered) {
1567 nl_table[unit].groups = groups;
1568 nl_table[unit].listeners = listeners;
1569 nl_table[unit].cb_mutex = cb_mutex;
1570 nl_table[unit].module = module;
1571 nl_table[unit].registered = 1;
Jesper Juhlf937f1f462007-10-15 01:39:12 -07001572 } else {
1573 kfree(listeners);
Denis V. Lunev869e58f2008-01-18 23:53:31 -08001574 nl_table[unit].registered++;
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001575 }
Patrick McHardy77247bb2005-08-14 19:27:13 -07001576 netlink_table_ungrab();
Harald Welte4fdb3bb2005-08-09 19:40:55 -07001577 return sk;
1578
Harald Welte4fdb3bb2005-08-09 19:40:55 -07001579out_sock_release:
Patrick McHardy4277a082006-03-20 18:52:01 -08001580 kfree(listeners);
Denis V. Lunev9dfbec12008-02-29 11:17:56 -08001581 netlink_kernel_release(sk);
Pavel Emelyanov23fe1862008-01-30 19:31:06 -08001582 return NULL;
1583
1584out_sock_release_nosk:
Harald Welte4fdb3bb2005-08-09 19:40:55 -07001585 sock_release(sock);
Patrick McHardy77247bb2005-08-14 19:27:13 -07001586 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001588EXPORT_SYMBOL(netlink_kernel_create);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589
Denis V. Lunevb7c6ba62008-01-28 14:41:19 -08001590
1591void
1592netlink_kernel_release(struct sock *sk)
1593{
Denis V. Lunevedf02082008-02-29 11:18:32 -08001594 sk_release_kernel(sk);
Denis V. Lunevb7c6ba62008-01-28 14:41:19 -08001595}
1596EXPORT_SYMBOL(netlink_kernel_release);
1597
1598
Johannes Berg6c04bb12009-07-10 09:51:32 +00001599static void netlink_free_old_listeners(struct rcu_head *rcu_head)
1600{
1601 struct listeners_rcu_head *lrh;
1602
1603 lrh = container_of(rcu_head, struct listeners_rcu_head, rcu_head);
1604 kfree(lrh->ptr);
1605}
1606
Johannes Bergd136f1b2009-09-12 03:03:15 +00001607int __netlink_change_ngroups(struct sock *sk, unsigned int groups)
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001608{
1609 unsigned long *listeners, *old = NULL;
Johannes Berg6c04bb12009-07-10 09:51:32 +00001610 struct listeners_rcu_head *old_rcu_head;
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001611 struct netlink_table *tbl = &nl_table[sk->sk_protocol];
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001612
1613 if (groups < 32)
1614 groups = 32;
1615
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001616 if (NLGRPSZ(tbl->groups) < NLGRPSZ(groups)) {
Johannes Berg6c04bb12009-07-10 09:51:32 +00001617 listeners = kzalloc(NLGRPSZ(groups) +
1618 sizeof(struct listeners_rcu_head),
1619 GFP_ATOMIC);
Johannes Bergd136f1b2009-09-12 03:03:15 +00001620 if (!listeners)
1621 return -ENOMEM;
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001622 old = tbl->listeners;
1623 memcpy(listeners, old, NLGRPSZ(tbl->groups));
1624 rcu_assign_pointer(tbl->listeners, listeners);
Johannes Berg6c04bb12009-07-10 09:51:32 +00001625 /*
1626 * Free the old memory after an RCU grace period so we
1627 * don't leak it. We use call_rcu() here in order to be
1628 * able to call this function from atomic contexts. The
1629 * allocation of this memory will have reserved enough
1630 * space for struct listeners_rcu_head at the end.
1631 */
1632 old_rcu_head = (void *)(tbl->listeners +
1633 NLGRPLONGS(tbl->groups));
1634 old_rcu_head->ptr = old;
1635 call_rcu(&old_rcu_head->rcu_head, netlink_free_old_listeners);
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001636 }
1637 tbl->groups = groups;
1638
Johannes Bergd136f1b2009-09-12 03:03:15 +00001639 return 0;
1640}
1641
1642/**
1643 * netlink_change_ngroups - change number of multicast groups
1644 *
1645 * This changes the number of multicast groups that are available
1646 * on a certain netlink family. Note that it is not possible to
1647 * change the number of groups to below 32. Also note that it does
1648 * not implicitly call netlink_clear_multicast_users() when the
1649 * number of groups is reduced.
1650 *
1651 * @sk: The kernel netlink socket, as returned by netlink_kernel_create().
1652 * @groups: The new number of groups.
1653 */
1654int netlink_change_ngroups(struct sock *sk, unsigned int groups)
1655{
1656 int err;
1657
1658 netlink_table_grab();
1659 err = __netlink_change_ngroups(sk, groups);
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001660 netlink_table_ungrab();
Johannes Bergd136f1b2009-09-12 03:03:15 +00001661
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001662 return err;
1663}
Johannes Bergb4ff4f02007-07-18 15:46:06 -07001664
Johannes Bergb8273572009-09-24 15:44:05 -07001665void __netlink_clear_multicast_users(struct sock *ksk, unsigned int group)
1666{
1667 struct sock *sk;
1668 struct hlist_node *node;
1669 struct netlink_table *tbl = &nl_table[ksk->sk_protocol];
1670
1671 sk_for_each_bound(sk, node, &tbl->mc_list)
1672 netlink_update_socket_mc(nlk_sk(sk), group, 0);
1673}
1674
Johannes Berg84659eb2007-07-18 15:47:05 -07001675/**
1676 * netlink_clear_multicast_users - kick off multicast listeners
1677 *
1678 * This function removes all listeners from the given group.
1679 * @ksk: The kernel netlink socket, as returned by
1680 * netlink_kernel_create().
1681 * @group: The multicast group to clear.
1682 */
1683void netlink_clear_multicast_users(struct sock *ksk, unsigned int group)
1684{
Johannes Berg84659eb2007-07-18 15:47:05 -07001685 netlink_table_grab();
Johannes Bergb8273572009-09-24 15:44:05 -07001686 __netlink_clear_multicast_users(ksk, group);
Johannes Berg84659eb2007-07-18 15:47:05 -07001687 netlink_table_ungrab();
1688}
Johannes Berg84659eb2007-07-18 15:47:05 -07001689
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690void netlink_set_nonroot(int protocol, unsigned int flags)
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09001691{
1692 if ((unsigned int)protocol < MAX_LINKS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 nl_table[protocol].nl_nonroot = flags;
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09001694}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001695EXPORT_SYMBOL(netlink_set_nonroot);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696
1697static void netlink_destroy_callback(struct netlink_callback *cb)
1698{
Wei Yongjun91744f62009-02-25 00:34:41 +00001699 kfree_skb(cb->skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 kfree(cb);
1701}
1702
1703/*
1704 * It looks a bit ugly.
1705 * It would be better to create kernel thread.
1706 */
1707
1708static int netlink_dump(struct sock *sk)
1709{
1710 struct netlink_sock *nlk = nlk_sk(sk);
1711 struct netlink_callback *cb;
1712 struct sk_buff *skb;
1713 struct nlmsghdr *nlh;
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001714 int len, err = -ENOBUFS;
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09001715
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 skb = sock_rmalloc(sk, NLMSG_GOODSIZE, 0, GFP_KERNEL);
1717 if (!skb)
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001718 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001720 mutex_lock(nlk->cb_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721
1722 cb = nlk->cb;
1723 if (cb == NULL) {
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001724 err = -EINVAL;
1725 goto errout_skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 }
1727
1728 len = cb->dump(skb, cb);
1729
1730 if (len > 0) {
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001731 mutex_unlock(nlk->cb_mutex);
Stephen Hemmingerb1153f22008-03-21 15:46:12 -07001732
1733 if (sk_filter(sk, skb))
1734 kfree_skb(skb);
1735 else {
1736 skb_queue_tail(&sk->sk_receive_queue, skb);
1737 sk->sk_data_ready(sk, skb->len);
1738 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 return 0;
1740 }
1741
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001742 nlh = nlmsg_put_answer(skb, cb, NLMSG_DONE, sizeof(len), NLM_F_MULTI);
1743 if (!nlh)
1744 goto errout_skb;
1745
1746 memcpy(nlmsg_data(nlh), &len, sizeof(len));
1747
Stephen Hemmingerb1153f22008-03-21 15:46:12 -07001748 if (sk_filter(sk, skb))
1749 kfree_skb(skb);
1750 else {
1751 skb_queue_tail(&sk->sk_receive_queue, skb);
1752 sk->sk_data_ready(sk, skb->len);
1753 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754
Thomas Grafa8f74b22005-11-10 02:25:52 +01001755 if (cb->done)
1756 cb->done(cb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 nlk->cb = NULL;
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001758 mutex_unlock(nlk->cb_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759
1760 netlink_destroy_callback(cb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 return 0;
Thomas Graf17977542005-06-18 22:53:48 -07001762
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001763errout_skb:
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001764 mutex_unlock(nlk->cb_mutex);
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001765 kfree_skb(skb);
1766errout:
1767 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768}
1769
1770int netlink_dump_start(struct sock *ssk, struct sk_buff *skb,
Patrick McHardy3a6c2b42009-08-25 16:07:40 +02001771 const struct nlmsghdr *nlh,
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001772 int (*dump)(struct sk_buff *skb,
1773 struct netlink_callback *),
1774 int (*done)(struct netlink_callback *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775{
1776 struct netlink_callback *cb;
1777 struct sock *sk;
1778 struct netlink_sock *nlk;
1779
Panagiotis Issaris0da974f2006-07-21 14:51:30 -07001780 cb = kzalloc(sizeof(*cb), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 if (cb == NULL)
1782 return -ENOBUFS;
1783
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 cb->dump = dump;
1785 cb->done = done;
1786 cb->nlh = nlh;
1787 atomic_inc(&skb->users);
1788 cb->skb = skb;
1789
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +09001790 sk = netlink_lookup(sock_net(ssk), ssk->sk_protocol, NETLINK_CB(skb).pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 if (sk == NULL) {
1792 netlink_destroy_callback(cb);
1793 return -ECONNREFUSED;
1794 }
1795 nlk = nlk_sk(sk);
Herbert Xu3f660d62007-05-03 03:17:14 -07001796 /* A dump is in progress... */
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001797 mutex_lock(nlk->cb_mutex);
Herbert Xu3f660d62007-05-03 03:17:14 -07001798 if (nlk->cb) {
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001799 mutex_unlock(nlk->cb_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 netlink_destroy_callback(cb);
1801 sock_put(sk);
1802 return -EBUSY;
1803 }
1804 nlk->cb = cb;
Patrick McHardyaf65bdf2007-04-20 14:14:21 -07001805 mutex_unlock(nlk->cb_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
1807 netlink_dump(sk);
1808 sock_put(sk);
Denis V. Lunev5c582982007-10-23 20:29:25 -07001809
1810 /* We successfully started a dump, by returning -EINTR we
1811 * signal not to send ACK even if it was requested.
1812 */
1813 return -EINTR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001815EXPORT_SYMBOL(netlink_dump_start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
1817void netlink_ack(struct sk_buff *in_skb, struct nlmsghdr *nlh, int err)
1818{
1819 struct sk_buff *skb;
1820 struct nlmsghdr *rep;
1821 struct nlmsgerr *errmsg;
Thomas Graf339bf982006-11-10 14:10:15 -08001822 size_t payload = sizeof(*errmsg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823
Thomas Graf339bf982006-11-10 14:10:15 -08001824 /* error messages get the original request appened */
1825 if (err)
1826 payload += nlmsg_len(nlh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827
Thomas Graf339bf982006-11-10 14:10:15 -08001828 skb = nlmsg_new(payload, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 if (!skb) {
1830 struct sock *sk;
1831
YOSHIFUJI Hideaki3b1e0a62008-03-26 02:26:21 +09001832 sk = netlink_lookup(sock_net(in_skb->sk),
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001833 in_skb->sk->sk_protocol,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 NETLINK_CB(in_skb).pid);
1835 if (sk) {
1836 sk->sk_err = ENOBUFS;
1837 sk->sk_error_report(sk);
1838 sock_put(sk);
1839 }
1840 return;
1841 }
1842
1843 rep = __nlmsg_put(skb, NETLINK_CB(in_skb).pid, nlh->nlmsg_seq,
John Fastabend5dba93a2009-09-25 13:11:44 +00001844 NLMSG_ERROR, payload, 0);
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001845 errmsg = nlmsg_data(rep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 errmsg->error = err;
Thomas Grafbf8b79e2006-08-04 23:03:29 -07001847 memcpy(&errmsg->msg, nlh, err ? nlh->nlmsg_len : sizeof(*nlh));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 netlink_unicast(in_skb->sk, skb, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
1849}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001850EXPORT_SYMBOL(netlink_ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -07001852int netlink_rcv_skb(struct sk_buff *skb, int (*cb)(struct sk_buff *,
Thomas Graf1d00a4e2007-03-22 23:30:12 -07001853 struct nlmsghdr *))
Thomas Graf82ace472005-11-10 02:25:53 +01001854{
Thomas Graf82ace472005-11-10 02:25:53 +01001855 struct nlmsghdr *nlh;
1856 int err;
1857
1858 while (skb->len >= nlmsg_total_size(0)) {
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -07001859 int msglen;
1860
Arnaldo Carvalho de Melob529ccf2007-04-25 19:08:35 -07001861 nlh = nlmsg_hdr(skb);
Thomas Grafd35b6852007-03-22 23:28:46 -07001862 err = 0;
Thomas Graf82ace472005-11-10 02:25:53 +01001863
Martin Murrayad8e4b72006-01-10 13:02:29 -08001864 if (nlh->nlmsg_len < NLMSG_HDRLEN || skb->len < nlh->nlmsg_len)
Thomas Graf82ace472005-11-10 02:25:53 +01001865 return 0;
1866
Thomas Grafd35b6852007-03-22 23:28:46 -07001867 /* Only requests are handled by the kernel */
1868 if (!(nlh->nlmsg_flags & NLM_F_REQUEST))
Denis V. Lunev5c582982007-10-23 20:29:25 -07001869 goto ack;
Thomas Grafd35b6852007-03-22 23:28:46 -07001870
Thomas Graf45e7ae72007-03-22 23:29:10 -07001871 /* Skip control messages */
1872 if (nlh->nlmsg_type < NLMSG_MIN_TYPE)
Denis V. Lunev5c582982007-10-23 20:29:25 -07001873 goto ack;
Thomas Graf45e7ae72007-03-22 23:29:10 -07001874
Thomas Graf1d00a4e2007-03-22 23:30:12 -07001875 err = cb(skb, nlh);
Denis V. Lunev5c582982007-10-23 20:29:25 -07001876 if (err == -EINTR)
1877 goto skip;
1878
1879ack:
Thomas Grafd35b6852007-03-22 23:28:46 -07001880 if (nlh->nlmsg_flags & NLM_F_ACK || err)
Thomas Graf82ace472005-11-10 02:25:53 +01001881 netlink_ack(skb, nlh, err);
Thomas Graf82ace472005-11-10 02:25:53 +01001882
Denis V. Lunev5c582982007-10-23 20:29:25 -07001883skip:
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001884 msglen = NLMSG_ALIGN(nlh->nlmsg_len);
Denis V. Lunevcd40b7d2007-10-10 21:15:29 -07001885 if (msglen > skb->len)
1886 msglen = skb->len;
1887 skb_pull(skb, msglen);
Thomas Graf82ace472005-11-10 02:25:53 +01001888 }
1889
1890 return 0;
1891}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001892EXPORT_SYMBOL(netlink_rcv_skb);
Thomas Graf82ace472005-11-10 02:25:53 +01001893
1894/**
Thomas Grafd387f6a2006-08-15 00:31:06 -07001895 * nlmsg_notify - send a notification netlink message
1896 * @sk: netlink socket to use
1897 * @skb: notification message
1898 * @pid: destination netlink pid for reports or 0
1899 * @group: destination multicast group or 0
1900 * @report: 1 to report back, 0 to disable
1901 * @flags: allocation flags
1902 */
1903int nlmsg_notify(struct sock *sk, struct sk_buff *skb, u32 pid,
1904 unsigned int group, int report, gfp_t flags)
1905{
1906 int err = 0;
1907
1908 if (group) {
1909 int exclude_pid = 0;
1910
1911 if (report) {
1912 atomic_inc(&skb->users);
1913 exclude_pid = pid;
1914 }
1915
Pablo Neira Ayuso1ce85fe2009-02-24 23:18:28 -08001916 /* errors reported via destination sk->sk_err, but propagate
1917 * delivery errors if NETLINK_BROADCAST_ERROR flag is set */
1918 err = nlmsg_multicast(sk, skb, exclude_pid, group, flags);
Thomas Grafd387f6a2006-08-15 00:31:06 -07001919 }
1920
Pablo Neira Ayuso1ce85fe2009-02-24 23:18:28 -08001921 if (report) {
1922 int err2;
1923
1924 err2 = nlmsg_unicast(sk, skb, pid);
1925 if (!err || err == -ESRCH)
1926 err = err2;
1927 }
Thomas Grafd387f6a2006-08-15 00:31:06 -07001928
1929 return err;
1930}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001931EXPORT_SYMBOL(nlmsg_notify);
Thomas Grafd387f6a2006-08-15 00:31:06 -07001932
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933#ifdef CONFIG_PROC_FS
1934struct nl_seq_iter {
Denis V. Luneve372c412007-11-19 22:31:54 -08001935 struct seq_net_private p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 int link;
1937 int hash_idx;
1938};
1939
1940static struct sock *netlink_seq_socket_idx(struct seq_file *seq, loff_t pos)
1941{
1942 struct nl_seq_iter *iter = seq->private;
1943 int i, j;
1944 struct sock *s;
1945 struct hlist_node *node;
1946 loff_t off = 0;
1947
Patrick McHardy6ac552f2007-12-04 00:19:38 -08001948 for (i = 0; i < MAX_LINKS; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 struct nl_pid_hash *hash = &nl_table[i].hash;
1950
1951 for (j = 0; j <= hash->mask; j++) {
1952 sk_for_each(s, node, &hash->table[j]) {
YOSHIFUJI Hideaki12188542008-03-26 02:36:06 +09001953 if (sock_net(s) != seq_file_net(seq))
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001954 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 if (off == pos) {
1956 iter->link = i;
1957 iter->hash_idx = j;
1958 return s;
1959 }
1960 ++off;
1961 }
1962 }
1963 }
1964 return NULL;
1965}
1966
1967static void *netlink_seq_start(struct seq_file *seq, loff_t *pos)
Eric Dumazet9a429c42008-01-01 21:58:02 -08001968 __acquires(nl_table_lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969{
1970 read_lock(&nl_table_lock);
1971 return *pos ? netlink_seq_socket_idx(seq, *pos - 1) : SEQ_START_TOKEN;
1972}
1973
1974static void *netlink_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1975{
1976 struct sock *s;
1977 struct nl_seq_iter *iter;
1978 int i, j;
1979
1980 ++*pos;
1981
1982 if (v == SEQ_START_TOKEN)
1983 return netlink_seq_socket_idx(seq, 0);
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09001984
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02001985 iter = seq->private;
1986 s = v;
1987 do {
1988 s = sk_next(s);
YOSHIFUJI Hideaki12188542008-03-26 02:36:06 +09001989 } while (s && sock_net(s) != seq_file_net(seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 if (s)
1991 return s;
1992
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 i = iter->link;
1994 j = iter->hash_idx + 1;
1995
1996 do {
1997 struct nl_pid_hash *hash = &nl_table[i].hash;
1998
1999 for (; j <= hash->mask; j++) {
2000 s = sk_head(&hash->table[j]);
YOSHIFUJI Hideaki12188542008-03-26 02:36:06 +09002001 while (s && sock_net(s) != seq_file_net(seq))
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02002002 s = sk_next(s);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 if (s) {
2004 iter->link = i;
2005 iter->hash_idx = j;
2006 return s;
2007 }
2008 }
2009
2010 j = 0;
2011 } while (++i < MAX_LINKS);
2012
2013 return NULL;
2014}
2015
2016static void netlink_seq_stop(struct seq_file *seq, void *v)
Eric Dumazet9a429c42008-01-01 21:58:02 -08002017 __releases(nl_table_lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018{
2019 read_unlock(&nl_table_lock);
2020}
2021
2022
2023static int netlink_seq_show(struct seq_file *seq, void *v)
2024{
2025 if (v == SEQ_START_TOKEN)
2026 seq_puts(seq,
2027 "sk Eth Pid Groups "
Masatake YAMATOcf0aa4e2010-02-27 19:45:37 +00002028 "Rmem Wmem Dump Locks Drops Inode\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 else {
2030 struct sock *s = v;
2031 struct netlink_sock *nlk = nlk_sk(s);
2032
Masatake YAMATOcf0aa4e2010-02-27 19:45:37 +00002033 seq_printf(seq, "%p %-3d %-6d %08x %-8d %-8d %p %-8d %-8d %-8lu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 s,
2035 s->sk_protocol,
2036 nlk->pid,
Patrick McHardy513c2502005-09-06 15:43:59 -07002037 nlk->groups ? (u32)nlk->groups[0] : 0,
Eric Dumazet31e6d362009-06-17 19:05:41 -07002038 sk_rmem_alloc_get(s),
2039 sk_wmem_alloc_get(s),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 nlk->cb,
Pablo Neira Ayuso38938bf2009-03-24 16:37:55 -07002041 atomic_read(&s->sk_refcnt),
Masatake YAMATOcf0aa4e2010-02-27 19:45:37 +00002042 atomic_read(&s->sk_drops),
2043 sock_i_ino(s)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 );
2045
2046 }
2047 return 0;
2048}
2049
Philippe De Muyter56b3d972007-07-10 23:07:31 -07002050static const struct seq_operations netlink_seq_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 .start = netlink_seq_start,
2052 .next = netlink_seq_next,
2053 .stop = netlink_seq_stop,
2054 .show = netlink_seq_show,
2055};
2056
2057
2058static int netlink_seq_open(struct inode *inode, struct file *file)
2059{
Denis V. Luneve372c412007-11-19 22:31:54 -08002060 return seq_open_net(inode, file, &netlink_seq_ops,
2061 sizeof(struct nl_seq_iter));
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02002062}
2063
Arjan van de Venda7071d2007-02-12 00:55:36 -08002064static const struct file_operations netlink_seq_fops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 .owner = THIS_MODULE,
2066 .open = netlink_seq_open,
2067 .read = seq_read,
2068 .llseek = seq_lseek,
Denis V. Luneve372c412007-11-19 22:31:54 -08002069 .release = seq_release_net,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070};
2071
2072#endif
2073
2074int netlink_register_notifier(struct notifier_block *nb)
2075{
Alan Sterne041c682006-03-27 01:16:30 -08002076 return atomic_notifier_chain_register(&netlink_chain, nb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08002078EXPORT_SYMBOL(netlink_register_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079
2080int netlink_unregister_notifier(struct notifier_block *nb)
2081{
Alan Sterne041c682006-03-27 01:16:30 -08002082 return atomic_notifier_chain_unregister(&netlink_chain, nb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083}
Patrick McHardy6ac552f2007-12-04 00:19:38 -08002084EXPORT_SYMBOL(netlink_unregister_notifier);
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09002085
Eric Dumazet90ddc4f2005-12-22 12:49:22 -08002086static const struct proto_ops netlink_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 .family = PF_NETLINK,
2088 .owner = THIS_MODULE,
2089 .release = netlink_release,
2090 .bind = netlink_bind,
2091 .connect = netlink_connect,
2092 .socketpair = sock_no_socketpair,
2093 .accept = sock_no_accept,
2094 .getname = netlink_getname,
2095 .poll = datagram_poll,
2096 .ioctl = sock_no_ioctl,
2097 .listen = sock_no_listen,
2098 .shutdown = sock_no_shutdown,
Patrick McHardy9a4595b2005-08-15 12:32:15 -07002099 .setsockopt = netlink_setsockopt,
2100 .getsockopt = netlink_getsockopt,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 .sendmsg = netlink_sendmsg,
2102 .recvmsg = netlink_recvmsg,
2103 .mmap = sock_no_mmap,
2104 .sendpage = sock_no_sendpage,
2105};
2106
Stephen Hemmingerec1b4cf2009-10-05 05:58:39 +00002107static const struct net_proto_family netlink_family_ops = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 .family = PF_NETLINK,
2109 .create = netlink_create,
2110 .owner = THIS_MODULE, /* for consistency 8) */
2111};
2112
Pavel Emelyanov46650792007-10-08 20:38:39 -07002113static int __net_init netlink_net_init(struct net *net)
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02002114{
2115#ifdef CONFIG_PROC_FS
2116 if (!proc_net_fops_create(net, "netlink", 0, &netlink_seq_fops))
2117 return -ENOMEM;
2118#endif
2119 return 0;
2120}
2121
Pavel Emelyanov46650792007-10-08 20:38:39 -07002122static void __net_exit netlink_net_exit(struct net *net)
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02002123{
2124#ifdef CONFIG_PROC_FS
2125 proc_net_remove(net, "netlink");
2126#endif
2127}
2128
Denis V. Lunev022cbae2007-11-13 03:23:50 -08002129static struct pernet_operations __net_initdata netlink_net_ops = {
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02002130 .init = netlink_net_init,
2131 .exit = netlink_net_exit,
2132};
2133
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134static int __init netlink_proto_init(void)
2135{
2136 struct sk_buff *dummy_skb;
2137 int i;
Denis Cheng26ff5dd2007-09-16 16:36:02 -07002138 unsigned long limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 unsigned int order;
2140 int err = proto_register(&netlink_proto, 0);
2141
2142 if (err != 0)
2143 goto out;
2144
YOSHIFUJI Hideakief047f52006-09-01 00:29:06 -07002145 BUILD_BUG_ON(sizeof(struct netlink_skb_parms) > sizeof(dummy_skb->cb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146
Panagiotis Issaris0da974f2006-07-21 14:51:30 -07002147 nl_table = kcalloc(MAX_LINKS, sizeof(*nl_table), GFP_KERNEL);
Akinobu Mitafab2caf2006-08-29 02:15:24 -07002148 if (!nl_table)
2149 goto panic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150
Jan Beulich44813742009-09-21 17:03:05 -07002151 if (totalram_pages >= (128 * 1024))
2152 limit = totalram_pages >> (21 - PAGE_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 else
Jan Beulich44813742009-09-21 17:03:05 -07002154 limit = totalram_pages >> (23 - PAGE_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155
Denis Cheng26ff5dd2007-09-16 16:36:02 -07002156 order = get_bitmask_order(limit) - 1 + PAGE_SHIFT;
2157 limit = (1UL << order) / sizeof(struct hlist_head);
2158 order = get_bitmask_order(min(limit, (unsigned long)UINT_MAX)) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
2160 for (i = 0; i < MAX_LINKS; i++) {
2161 struct nl_pid_hash *hash = &nl_table[i].hash;
2162
Eric Dumazetea729122007-12-11 02:09:47 -08002163 hash->table = nl_pid_hash_zalloc(1 * sizeof(*hash->table));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 if (!hash->table) {
2165 while (i-- > 0)
2166 nl_pid_hash_free(nl_table[i].hash.table,
2167 1 * sizeof(*hash->table));
2168 kfree(nl_table);
Akinobu Mitafab2caf2006-08-29 02:15:24 -07002169 goto panic;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 hash->max_shift = order;
2172 hash->shift = 0;
2173 hash->mask = 0;
2174 hash->rehash_time = jiffies;
2175 }
2176
2177 sock_register(&netlink_family_ops);
Eric W. Biedermanb4b51022007-09-12 13:05:38 +02002178 register_pernet_subsys(&netlink_net_ops);
YOSHIFUJI Hideaki746fac42007-02-09 23:25:07 +09002179 /* The netlink device handler may be needed early. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 rtnetlink_init();
2181out:
2182 return err;
Akinobu Mitafab2caf2006-08-29 02:15:24 -07002183panic:
2184 panic("netlink_init: Cannot allocate nl_table\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185}
2186
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187core_initcall(netlink_proto_init);