1 /*
2  *  Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
3  *
4  *  This program is free software; you can redistribute it and/or modify
5  *  it under the terms of the GNU General Public License as published by
6  *  the Free Software Foundation; either version 2, or (at your option)
7  *  any later version.
8  *
9  *  This program is distributed in the hope that it will be useful,
10  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
11  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  *  GNU General Public License for more details.
13  *
14  *  You should have received a copy of the GNU General Public License
15  *  along with this program; see the file COPYING.  If not, write to
16  *  the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
17  */
18 
19 /*
20  * fsnotify inode mark locking/lifetime/and refcnting
21  *
22  * REFCNT:
23  * The mark->refcnt tells how many "things" in the kernel currently are
24  * referencing this object.  The object typically will live inside the kernel
25  * with a refcnt of 2, one for each list it is on (i_list, g_list).  Any task
26  * which can find this object holding the appropriete locks, can take a reference
27  * and the object itself is guaranteed to survive until the reference is dropped.
28  *
29  * LOCKING:
30  * There are 3 spinlocks involved with fsnotify inode marks and they MUST
31  * be taken in order as follows:
32  *
33  * mark->lock
34  * group->mark_lock
35  * inode->i_lock
36  *
37  * mark->lock protects 2 things, mark->group and mark->inode.  You must hold
38  * that lock to dereference either of these things (they could be NULL even with
39  * the lock)
40  *
41  * group->mark_lock protects the marks_list anchored inside a given group
42  * and each mark is hooked via the g_list.  It also sorta protects the
43  * free_g_list, which when used is anchored by a private list on the stack of the
44  * task which held the group->mark_lock.
45  *
46  * inode->i_lock protects the i_fsnotify_marks list anchored inside a
47  * given inode and each mark is hooked via the i_list. (and sorta the
48  * free_i_list)
49  *
50  *
51  * LIFETIME:
52  * Inode marks survive between when they are added to an inode and when their
53  * refcnt==0.
54  *
55  * The inode mark can be cleared for a number of different reasons including:
56  * - The inode is unlinked for the last time.  (fsnotify_inode_remove)
57  * - The inode is being evicted from cache. (fsnotify_inode_delete)
58  * - The fs the inode is on is unmounted.  (fsnotify_inode_delete/fsnotify_unmount_inodes)
59  * - Something explicitly requests that it be removed.  (fsnotify_destroy_mark)
60  * - The fsnotify_group associated with the mark is going away and all such marks
61  *   need to be cleaned up. (fsnotify_clear_marks_by_group)
62  *
63  * Worst case we are given an inode and need to clean up all the marks on that
64  * inode.  We take i_lock and walk the i_fsnotify_marks safely.  For each
65  * mark on the list we take a reference (so the mark can't disappear under us).
66  * We remove that mark form the inode's list of marks and we add this mark to a
67  * private list anchored on the stack using i_free_list;  At this point we no
68  * longer fear anything finding the mark using the inode's list of marks.
69  *
70  * We can safely and locklessly run the private list on the stack of everything
71  * we just unattached from the original inode.  For each mark on the private list
72  * we grab the mark-> and can thus dereference mark->group and mark->inode.  If
73  * we see the group and inode are not NULL we take those locks.  Now holding all
74  * 3 locks we can completely remove the mark from other tasks finding it in the
75  * future.  Remember, 10 things might already be referencing this mark, but they
76  * better be holding a ref.  We drop our reference we took before we unhooked it
77  * from the inode.  When the ref hits 0 we can free the mark.
78  *
79  * Very similarly for freeing by group, except we use free_g_list.
80  *
81  * This has the very interesting property of being able to run concurrently with
82  * any (or all) other directions.
83  */
84 
85 #include <linux/fs.h>
86 #include <linux/init.h>
87 #include <linux/kernel.h>
88 #include <linux/kthread.h>
89 #include <linux/module.h>
90 #include <linux/mutex.h>
91 #include <linux/slab.h>
92 #include <linux/spinlock.h>
93 #include <linux/srcu.h>
94 
95 #include <linux/atomic.h>
96 
97 #include <linux/fsnotify_backend.h>
98 #include "fsnotify.h"
99 
100 struct srcu_struct fsnotify_mark_srcu;
101 static DEFINE_SPINLOCK(destroy_lock);
102 static LIST_HEAD(destroy_list);
103 static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
104 
fsnotify_get_mark(struct fsnotify_mark * mark)105 void fsnotify_get_mark(struct fsnotify_mark *mark)
106 {
107 	atomic_inc(&mark->refcnt);
108 }
109 
fsnotify_put_mark(struct fsnotify_mark * mark)110 void fsnotify_put_mark(struct fsnotify_mark *mark)
111 {
112 	if (atomic_dec_and_test(&mark->refcnt))
113 		mark->free_mark(mark);
114 }
115 
116 /*
117  * Any time a mark is getting freed we end up here.
118  * The caller had better be holding a reference to this mark so we don't actually
119  * do the final put under the mark->lock
120  */
fsnotify_destroy_mark(struct fsnotify_mark * mark)121 void fsnotify_destroy_mark(struct fsnotify_mark *mark)
122 {
123 	struct fsnotify_group *group;
124 	struct inode *inode = NULL;
125 
126 	spin_lock(&mark->lock);
127 
128 	group = mark->group;
129 
130 	/* something else already called this function on this mark */
131 	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
132 		spin_unlock(&mark->lock);
133 		return;
134 	}
135 
136 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
137 
138 	spin_lock(&group->mark_lock);
139 
140 	if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
141 		inode = mark->i.inode;
142 		fsnotify_destroy_inode_mark(mark);
143 	} else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
144 		fsnotify_destroy_vfsmount_mark(mark);
145 	else
146 		BUG();
147 
148 	list_del_init(&mark->g_list);
149 
150 	spin_unlock(&group->mark_lock);
151 	spin_unlock(&mark->lock);
152 
153 	spin_lock(&destroy_lock);
154 	list_add(&mark->destroy_list, &destroy_list);
155 	spin_unlock(&destroy_lock);
156 	wake_up(&destroy_waitq);
157 
158 	/*
159 	 * Some groups like to know that marks are being freed.  This is a
160 	 * callback to the group function to let it know that this mark
161 	 * is being freed.
162 	 */
163 	if (group->ops->freeing_mark)
164 		group->ops->freeing_mark(mark, group);
165 
166 	/*
167 	 * __fsnotify_update_child_dentry_flags(inode);
168 	 *
169 	 * I really want to call that, but we can't, we have no idea if the inode
170 	 * still exists the second we drop the mark->lock.
171 	 *
172 	 * The next time an event arrive to this inode from one of it's children
173 	 * __fsnotify_parent will see that the inode doesn't care about it's
174 	 * children and will update all of these flags then.  So really this
175 	 * is just a lazy update (and could be a perf win...)
176 	 */
177 
178 	if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
179 		iput(inode);
180 
181 	/*
182 	 * We don't necessarily have a ref on mark from caller so the above iput
183 	 * may have already destroyed it.  Don't touch from now on.
184 	 */
185 
186 	/*
187 	 * it's possible that this group tried to destroy itself, but this
188 	 * this mark was simultaneously being freed by inode.  If that's the
189 	 * case, we finish freeing the group here.
190 	 */
191 	if (unlikely(atomic_dec_and_test(&group->num_marks)))
192 		fsnotify_final_destroy_group(group);
193 }
194 
fsnotify_set_mark_mask_locked(struct fsnotify_mark * mark,__u32 mask)195 void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
196 {
197 	assert_spin_locked(&mark->lock);
198 
199 	mark->mask = mask;
200 
201 	if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
202 		fsnotify_set_inode_mark_mask_locked(mark, mask);
203 }
204 
fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark * mark,__u32 mask)205 void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
206 {
207 	assert_spin_locked(&mark->lock);
208 
209 	mark->ignored_mask = mask;
210 }
211 
212 /*
213  * Attach an initialized mark to a given group and fs object.
214  * These marks may be used for the fsnotify backend to determine which
215  * event types should be delivered to which group.
216  */
fsnotify_add_mark(struct fsnotify_mark * mark,struct fsnotify_group * group,struct inode * inode,struct vfsmount * mnt,int allow_dups)217 int fsnotify_add_mark(struct fsnotify_mark *mark,
218 		      struct fsnotify_group *group, struct inode *inode,
219 		      struct vfsmount *mnt, int allow_dups)
220 {
221 	int ret = 0;
222 
223 	BUG_ON(inode && mnt);
224 	BUG_ON(!inode && !mnt);
225 
226 	/*
227 	 * LOCKING ORDER!!!!
228 	 * mark->lock
229 	 * group->mark_lock
230 	 * inode->i_lock
231 	 */
232 	spin_lock(&mark->lock);
233 	spin_lock(&group->mark_lock);
234 
235 	mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE;
236 
237 	mark->group = group;
238 	list_add(&mark->g_list, &group->marks_list);
239 	atomic_inc(&group->num_marks);
240 	fsnotify_get_mark(mark); /* for i_list and g_list */
241 
242 	if (inode) {
243 		ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
244 		if (ret)
245 			goto err;
246 	} else if (mnt) {
247 		ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
248 		if (ret)
249 			goto err;
250 	} else {
251 		BUG();
252 	}
253 
254 	spin_unlock(&group->mark_lock);
255 
256 	/* this will pin the object if appropriate */
257 	fsnotify_set_mark_mask_locked(mark, mark->mask);
258 
259 	spin_unlock(&mark->lock);
260 
261 	if (inode)
262 		__fsnotify_update_child_dentry_flags(inode);
263 
264 	return ret;
265 err:
266 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
267 	list_del_init(&mark->g_list);
268 	mark->group = NULL;
269 	atomic_dec(&group->num_marks);
270 
271 	spin_unlock(&group->mark_lock);
272 	spin_unlock(&mark->lock);
273 
274 	spin_lock(&destroy_lock);
275 	list_add(&mark->destroy_list, &destroy_list);
276 	spin_unlock(&destroy_lock);
277 	wake_up(&destroy_waitq);
278 
279 	return ret;
280 }
281 
282 /*
283  * clear any marks in a group in which mark->flags & flags is true
284  */
fsnotify_clear_marks_by_group_flags(struct fsnotify_group * group,unsigned int flags)285 void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
286 					 unsigned int flags)
287 {
288 	struct fsnotify_mark *lmark, *mark;
289 	LIST_HEAD(free_list);
290 
291 	spin_lock(&group->mark_lock);
292 	list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
293 		if (mark->flags & flags) {
294 			list_add(&mark->free_g_list, &free_list);
295 			list_del_init(&mark->g_list);
296 			fsnotify_get_mark(mark);
297 		}
298 	}
299 	spin_unlock(&group->mark_lock);
300 
301 	list_for_each_entry_safe(mark, lmark, &free_list, free_g_list) {
302 		fsnotify_destroy_mark(mark);
303 		fsnotify_put_mark(mark);
304 	}
305 }
306 
307 /*
308  * Given a group, destroy all of the marks associated with that group.
309  */
fsnotify_clear_marks_by_group(struct fsnotify_group * group)310 void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
311 {
312 	fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
313 }
314 
fsnotify_duplicate_mark(struct fsnotify_mark * new,struct fsnotify_mark * old)315 void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
316 {
317 	assert_spin_locked(&old->lock);
318 	new->i.inode = old->i.inode;
319 	new->m.mnt = old->m.mnt;
320 	new->group = old->group;
321 	new->mask = old->mask;
322 	new->free_mark = old->free_mark;
323 }
324 
325 /*
326  * Nothing fancy, just initialize lists and locks and counters.
327  */
fsnotify_init_mark(struct fsnotify_mark * mark,void (* free_mark)(struct fsnotify_mark * mark))328 void fsnotify_init_mark(struct fsnotify_mark *mark,
329 			void (*free_mark)(struct fsnotify_mark *mark))
330 {
331 	memset(mark, 0, sizeof(*mark));
332 	spin_lock_init(&mark->lock);
333 	atomic_set(&mark->refcnt, 1);
334 	mark->free_mark = free_mark;
335 }
336 
fsnotify_mark_destroy(void * ignored)337 static int fsnotify_mark_destroy(void *ignored)
338 {
339 	struct fsnotify_mark *mark, *next;
340 	LIST_HEAD(private_destroy_list);
341 
342 	for (;;) {
343 		spin_lock(&destroy_lock);
344 		/* exchange the list head */
345 		list_replace_init(&destroy_list, &private_destroy_list);
346 		spin_unlock(&destroy_lock);
347 
348 		synchronize_srcu(&fsnotify_mark_srcu);
349 
350 		list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) {
351 			list_del_init(&mark->destroy_list);
352 			fsnotify_put_mark(mark);
353 		}
354 
355 		wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
356 	}
357 
358 	return 0;
359 }
360 
fsnotify_mark_init(void)361 static int __init fsnotify_mark_init(void)
362 {
363 	struct task_struct *thread;
364 
365 	thread = kthread_run(fsnotify_mark_destroy, NULL,
366 			     "fsnotify_mark");
367 	if (IS_ERR(thread))
368 		panic("unable to start fsnotify mark destruction thread.");
369 
370 	return 0;
371 }
372 device_initcall(fsnotify_mark_init);
373