]> Pileus Git - ~andy/linux/blob - net/core/net_namespace.c
Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[~andy/linux] / net / core / net_namespace.c
1 #include <linux/workqueue.h>
2 #include <linux/rtnetlink.h>
3 #include <linux/cache.h>
4 #include <linux/slab.h>
5 #include <linux/list.h>
6 #include <linux/delay.h>
7 #include <linux/sched.h>
8 #include <linux/idr.h>
9 #include <linux/rculist.h>
10 #include <linux/nsproxy.h>
11 #include <linux/netdevice.h>
12 #include <net/net_namespace.h>
13 #include <net/netns/generic.h>
14 #include <net/rtnetlink.h>
15
16 /*
17  *      Our network namespace constructor/destructor lists
18  */
19
20 static LIST_HEAD(pernet_list);
21 static struct list_head *first_device = &pernet_list;
22 static DEFINE_MUTEX(net_mutex);
23
24 LIST_HEAD(net_namespace_list);
25 EXPORT_SYMBOL_GPL(net_namespace_list);
26
27 struct net init_net;
28 EXPORT_SYMBOL(init_net);
29
30 #define INITIAL_NET_GEN_PTRS    13 /* +1 for len +2 for rcu_head */
31
32 static void unregister_netdevices(struct net *net, struct list_head *list)
33 {
34         struct net_device *dev;
35         /* At exit all network devices most be removed from a network
36          * namespace.  Do this in the reverse order of registeration.
37          */
38         for_each_netdev_reverse(net, dev) {
39                 if (dev->rtnl_link_ops)
40                         dev->rtnl_link_ops->dellink(dev, list);
41                 else
42                         unregister_netdevice_queue(dev, list);
43         }
44 }
45
46 static int ops_init(const struct pernet_operations *ops, struct net *net)
47 {
48         int err;
49         if (ops->id && ops->size) {
50                 void *data = kzalloc(ops->size, GFP_KERNEL);
51                 if (!data)
52                         return -ENOMEM;
53
54                 err = net_assign_generic(net, *ops->id, data);
55                 if (err) {
56                         kfree(data);
57                         return err;
58                 }
59         }
60         if (ops->init)
61                 return ops->init(net);
62         return 0;
63 }
64
65 static void ops_free(const struct pernet_operations *ops, struct net *net)
66 {
67         if (ops->id && ops->size) {
68                 int id = *ops->id;
69                 kfree(net_generic(net, id));
70         }
71 }
72
73 /*
74  * setup_net runs the initializers for the network namespace object.
75  */
76 static __net_init int setup_net(struct net *net)
77 {
78         /* Must be called with net_mutex held */
79         const struct pernet_operations *ops, *saved_ops;
80         int error = 0;
81
82         atomic_set(&net->count, 1);
83
84 #ifdef NETNS_REFCNT_DEBUG
85         atomic_set(&net->use_count, 0);
86 #endif
87
88         list_for_each_entry(ops, &pernet_list, list) {
89                 error = ops_init(ops, net);
90                 if (error < 0)
91                         goto out_undo;
92         }
93 out:
94         return error;
95
96 out_undo:
97         /* Walk through the list backwards calling the exit functions
98          * for the pernet modules whose init functions did not fail.
99          */
100         saved_ops = ops;
101         list_for_each_entry_continue_reverse(ops, &pernet_list, list) {
102                 if (ops->exit)
103                         ops->exit(net);
104                 if (&ops->list == first_device) {
105                         LIST_HEAD(dev_kill_list);
106                         rtnl_lock();
107                         unregister_netdevices(net, &dev_kill_list);
108                         unregister_netdevice_many(&dev_kill_list);
109                         rtnl_unlock();
110                 }
111         }
112         ops = saved_ops;
113         list_for_each_entry_continue_reverse(ops, &pernet_list, list)
114                 ops_free(ops, net);
115
116         rcu_barrier();
117         goto out;
118 }
119
120 static struct net_generic *net_alloc_generic(void)
121 {
122         struct net_generic *ng;
123         size_t generic_size = sizeof(struct net_generic) +
124                 INITIAL_NET_GEN_PTRS * sizeof(void *);
125
126         ng = kzalloc(generic_size, GFP_KERNEL);
127         if (ng)
128                 ng->len = INITIAL_NET_GEN_PTRS;
129
130         return ng;
131 }
132
133 #ifdef CONFIG_NET_NS
134 static struct kmem_cache *net_cachep;
135 static struct workqueue_struct *netns_wq;
136
137 static struct net *net_alloc(void)
138 {
139         struct net *net = NULL;
140         struct net_generic *ng;
141
142         ng = net_alloc_generic();
143         if (!ng)
144                 goto out;
145
146         net = kmem_cache_zalloc(net_cachep, GFP_KERNEL);
147         if (!net)
148                 goto out_free;
149
150         rcu_assign_pointer(net->gen, ng);
151 out:
152         return net;
153
154 out_free:
155         kfree(ng);
156         goto out;
157 }
158
159 static void net_free(struct net *net)
160 {
161 #ifdef NETNS_REFCNT_DEBUG
162         if (unlikely(atomic_read(&net->use_count) != 0)) {
163                 printk(KERN_EMERG "network namespace not free! Usage: %d\n",
164                         atomic_read(&net->use_count));
165                 return;
166         }
167 #endif
168         kfree(net->gen);
169         kmem_cache_free(net_cachep, net);
170 }
171
172 static struct net *net_create(void)
173 {
174         struct net *net;
175         int rv;
176
177         net = net_alloc();
178         if (!net)
179                 return ERR_PTR(-ENOMEM);
180         mutex_lock(&net_mutex);
181         rv = setup_net(net);
182         if (rv == 0) {
183                 rtnl_lock();
184                 list_add_tail_rcu(&net->list, &net_namespace_list);
185                 rtnl_unlock();
186         }
187         mutex_unlock(&net_mutex);
188         if (rv < 0) {
189                 net_free(net);
190                 return ERR_PTR(rv);
191         }
192         return net;
193 }
194
195 struct net *copy_net_ns(unsigned long flags, struct net *old_net)
196 {
197         if (!(flags & CLONE_NEWNET))
198                 return get_net(old_net);
199         return net_create();
200 }
201
202 static DEFINE_SPINLOCK(cleanup_list_lock);
203 static LIST_HEAD(cleanup_list);  /* Must hold cleanup_list_lock to touch */
204
205 static void cleanup_net(struct work_struct *work)
206 {
207         const struct pernet_operations *ops;
208         struct net *net, *tmp;
209         LIST_HEAD(net_kill_list);
210
211         /* Atomically snapshot the list of namespaces to cleanup */
212         spin_lock_irq(&cleanup_list_lock);
213         list_replace_init(&cleanup_list, &net_kill_list);
214         spin_unlock_irq(&cleanup_list_lock);
215
216         mutex_lock(&net_mutex);
217
218         /* Don't let anyone else find us. */
219         rtnl_lock();
220         list_for_each_entry(net, &net_kill_list, cleanup_list)
221                 list_del_rcu(&net->list);
222         rtnl_unlock();
223
224         /*
225          * Another CPU might be rcu-iterating the list, wait for it.
226          * This needs to be before calling the exit() notifiers, so
227          * the rcu_barrier() below isn't sufficient alone.
228          */
229         synchronize_rcu();
230
231         /* Run all of the network namespace exit methods */
232         list_for_each_entry_reverse(ops, &pernet_list, list) {
233                 if (ops->exit) {
234                         list_for_each_entry(net, &net_kill_list, cleanup_list)
235                                 ops->exit(net);
236                 }
237                 if (&ops->list == first_device) {
238                         LIST_HEAD(dev_kill_list);
239                         rtnl_lock();
240                         list_for_each_entry(net, &net_kill_list, cleanup_list)
241                                 unregister_netdevices(net, &dev_kill_list);
242                         unregister_netdevice_many(&dev_kill_list);
243                         rtnl_unlock();
244                 }
245         }
246         /* Free the net generic variables */
247         list_for_each_entry_reverse(ops, &pernet_list, list) {
248                 if (ops->size && ops->id) {
249                         list_for_each_entry(net, &net_kill_list, cleanup_list)
250                                 ops_free(ops, net);
251                 }
252         }
253
254         mutex_unlock(&net_mutex);
255
256         /* Ensure there are no outstanding rcu callbacks using this
257          * network namespace.
258          */
259         rcu_barrier();
260
261         /* Finally it is safe to free my network namespace structure */
262         list_for_each_entry_safe(net, tmp, &net_kill_list, cleanup_list) {
263                 list_del_init(&net->cleanup_list);
264                 net_free(net);
265         }
266 }
267 static DECLARE_WORK(net_cleanup_work, cleanup_net);
268
269 void __put_net(struct net *net)
270 {
271         /* Cleanup the network namespace in process context */
272         unsigned long flags;
273
274         spin_lock_irqsave(&cleanup_list_lock, flags);
275         list_add(&net->cleanup_list, &cleanup_list);
276         spin_unlock_irqrestore(&cleanup_list_lock, flags);
277
278         queue_work(netns_wq, &net_cleanup_work);
279 }
280 EXPORT_SYMBOL_GPL(__put_net);
281
282 #else
283 struct net *copy_net_ns(unsigned long flags, struct net *old_net)
284 {
285         if (flags & CLONE_NEWNET)
286                 return ERR_PTR(-EINVAL);
287         return old_net;
288 }
289 #endif
290
291 struct net *get_net_ns_by_pid(pid_t pid)
292 {
293         struct task_struct *tsk;
294         struct net *net;
295
296         /* Lookup the network namespace */
297         net = ERR_PTR(-ESRCH);
298         rcu_read_lock();
299         tsk = find_task_by_vpid(pid);
300         if (tsk) {
301                 struct nsproxy *nsproxy;
302                 nsproxy = task_nsproxy(tsk);
303                 if (nsproxy)
304                         net = get_net(nsproxy->net_ns);
305         }
306         rcu_read_unlock();
307         return net;
308 }
309 EXPORT_SYMBOL_GPL(get_net_ns_by_pid);
310
311 static int __init net_ns_init(void)
312 {
313         struct net_generic *ng;
314
315 #ifdef CONFIG_NET_NS
316         net_cachep = kmem_cache_create("net_namespace", sizeof(struct net),
317                                         SMP_CACHE_BYTES,
318                                         SLAB_PANIC, NULL);
319
320         /* Create workqueue for cleanup */
321         netns_wq = create_singlethread_workqueue("netns");
322         if (!netns_wq)
323                 panic("Could not create netns workq");
324 #endif
325
326         ng = net_alloc_generic();
327         if (!ng)
328                 panic("Could not allocate generic netns");
329
330         rcu_assign_pointer(init_net.gen, ng);
331
332         mutex_lock(&net_mutex);
333         if (setup_net(&init_net))
334                 panic("Could not setup the initial network namespace");
335
336         rtnl_lock();
337         list_add_tail_rcu(&init_net.list, &net_namespace_list);
338         rtnl_unlock();
339
340         mutex_unlock(&net_mutex);
341
342         return 0;
343 }
344
345 pure_initcall(net_ns_init);
346
347 #ifdef CONFIG_NET_NS
348 static int __register_pernet_operations(struct list_head *list,
349                                         struct pernet_operations *ops)
350 {
351         struct net *net, *undo_net;
352         int error;
353
354         list_add_tail(&ops->list, list);
355         if (ops->init || (ops->id && ops->size)) {
356                 for_each_net(net) {
357                         error = ops_init(ops, net);
358                         if (error)
359                                 goto out_undo;
360                 }
361         }
362         return 0;
363
364 out_undo:
365         /* If I have an error cleanup all namespaces I initialized */
366         list_del(&ops->list);
367         if (ops->exit) {
368                 for_each_net(undo_net) {
369                         if (net_eq(undo_net, net))
370                                 goto undone;
371                         ops->exit(undo_net);
372                 }
373         }
374 undone:
375         if (ops->size && ops->id) {
376                 for_each_net(undo_net) {
377                         if (net_eq(undo_net, net))
378                                 goto freed;
379                         ops_free(ops, undo_net);
380                 }
381         }
382 freed:
383         return error;
384 }
385
386 static void __unregister_pernet_operations(struct pernet_operations *ops)
387 {
388         struct net *net;
389
390         list_del(&ops->list);
391         if (ops->exit)
392                 for_each_net(net)
393                         ops->exit(net);
394         if (ops->id && ops->size)
395                 for_each_net(net)
396                         ops_free(ops, net);
397 }
398
399 #else
400
401 static int __register_pernet_operations(struct list_head *list,
402                                         struct pernet_operations *ops)
403 {
404         int err = 0;
405         err = ops_init(ops, &init_net);
406         if (err)
407                 ops_free(ops, &init_net);
408         return err;
409         
410 }
411
412 static void __unregister_pernet_operations(struct pernet_operations *ops)
413 {
414         if (ops->exit)
415                 ops->exit(&init_net);
416         ops_free(ops, &init_net);
417 }
418
419 #endif /* CONFIG_NET_NS */
420
421 static DEFINE_IDA(net_generic_ids);
422
423 static int register_pernet_operations(struct list_head *list,
424                                       struct pernet_operations *ops)
425 {
426         int error;
427
428         if (ops->id) {
429 again:
430                 error = ida_get_new_above(&net_generic_ids, 1, ops->id);
431                 if (error < 0) {
432                         if (error == -EAGAIN) {
433                                 ida_pre_get(&net_generic_ids, GFP_KERNEL);
434                                 goto again;
435                         }
436                         return error;
437                 }
438         }
439         error = __register_pernet_operations(list, ops);
440         if (error && ops->id)
441                 ida_remove(&net_generic_ids, *ops->id);
442
443         return error;
444 }
445
446 static void unregister_pernet_operations(struct pernet_operations *ops)
447 {
448         
449         __unregister_pernet_operations(ops);
450         if (ops->id)
451                 ida_remove(&net_generic_ids, *ops->id);
452 }
453
454 /**
455  *      register_pernet_subsys - register a network namespace subsystem
456  *      @ops:  pernet operations structure for the subsystem
457  *
458  *      Register a subsystem which has init and exit functions
459  *      that are called when network namespaces are created and
460  *      destroyed respectively.
461  *
462  *      When registered all network namespace init functions are
463  *      called for every existing network namespace.  Allowing kernel
464  *      modules to have a race free view of the set of network namespaces.
465  *
466  *      When a new network namespace is created all of the init
467  *      methods are called in the order in which they were registered.
468  *
469  *      When a network namespace is destroyed all of the exit methods
470  *      are called in the reverse of the order with which they were
471  *      registered.
472  */
473 int register_pernet_subsys(struct pernet_operations *ops)
474 {
475         int error;
476         mutex_lock(&net_mutex);
477         error =  register_pernet_operations(first_device, ops);
478         mutex_unlock(&net_mutex);
479         return error;
480 }
481 EXPORT_SYMBOL_GPL(register_pernet_subsys);
482
483 /**
484  *      unregister_pernet_subsys - unregister a network namespace subsystem
485  *      @ops: pernet operations structure to manipulate
486  *
487  *      Remove the pernet operations structure from the list to be
488  *      used when network namespaces are created or destroyed.  In
489  *      addition run the exit method for all existing network
490  *      namespaces.
491  */
492 void unregister_pernet_subsys(struct pernet_operations *module)
493 {
494         mutex_lock(&net_mutex);
495         unregister_pernet_operations(module);
496         mutex_unlock(&net_mutex);
497 }
498 EXPORT_SYMBOL_GPL(unregister_pernet_subsys);
499
500 /**
501  *      register_pernet_device - register a network namespace device
502  *      @ops:  pernet operations structure for the subsystem
503  *
504  *      Register a device which has init and exit functions
505  *      that are called when network namespaces are created and
506  *      destroyed respectively.
507  *
508  *      When registered all network namespace init functions are
509  *      called for every existing network namespace.  Allowing kernel
510  *      modules to have a race free view of the set of network namespaces.
511  *
512  *      When a new network namespace is created all of the init
513  *      methods are called in the order in which they were registered.
514  *
515  *      When a network namespace is destroyed all of the exit methods
516  *      are called in the reverse of the order with which they were
517  *      registered.
518  */
519 int register_pernet_device(struct pernet_operations *ops)
520 {
521         int error;
522         mutex_lock(&net_mutex);
523         error = register_pernet_operations(&pernet_list, ops);
524         if (!error && (first_device == &pernet_list))
525                 first_device = &ops->list;
526         mutex_unlock(&net_mutex);
527         return error;
528 }
529 EXPORT_SYMBOL_GPL(register_pernet_device);
530
531 /**
532  *      unregister_pernet_device - unregister a network namespace netdevice
533  *      @ops: pernet operations structure to manipulate
534  *
535  *      Remove the pernet operations structure from the list to be
536  *      used when network namespaces are created or destroyed.  In
537  *      addition run the exit method for all existing network
538  *      namespaces.
539  */
540 void unregister_pernet_device(struct pernet_operations *ops)
541 {
542         mutex_lock(&net_mutex);
543         if (&ops->list == first_device)
544                 first_device = first_device->next;
545         unregister_pernet_operations(ops);
546         mutex_unlock(&net_mutex);
547 }
548 EXPORT_SYMBOL_GPL(unregister_pernet_device);
549
550 static void net_generic_release(struct rcu_head *rcu)
551 {
552         struct net_generic *ng;
553
554         ng = container_of(rcu, struct net_generic, rcu);
555         kfree(ng);
556 }
557
558 int net_assign_generic(struct net *net, int id, void *data)
559 {
560         struct net_generic *ng, *old_ng;
561
562         BUG_ON(!mutex_is_locked(&net_mutex));
563         BUG_ON(id == 0);
564
565         ng = old_ng = net->gen;
566         if (old_ng->len >= id)
567                 goto assign;
568
569         ng = kzalloc(sizeof(struct net_generic) +
570                         id * sizeof(void *), GFP_KERNEL);
571         if (ng == NULL)
572                 return -ENOMEM;
573
574         /*
575          * Some synchronisation notes:
576          *
577          * The net_generic explores the net->gen array inside rcu
578          * read section. Besides once set the net->gen->ptr[x]
579          * pointer never changes (see rules in netns/generic.h).
580          *
581          * That said, we simply duplicate this array and schedule
582          * the old copy for kfree after a grace period.
583          */
584
585         ng->len = id;
586         memcpy(&ng->ptr, &old_ng->ptr, old_ng->len * sizeof(void*));
587
588         rcu_assign_pointer(net->gen, ng);
589         call_rcu(&old_ng->rcu, net_generic_release);
590 assign:
591         ng->ptr[id - 1] = data;
592         return 0;
593 }
594 EXPORT_SYMBOL_GPL(net_assign_generic);