xref: /linux/fs/notify/mark.c (revision cc04a46f11ea046ed53e2c832ae29e4790f7e35f)
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 group->recnt and mark->refcnt tell how many "things" in the kernel
24  * currently are referencing the objects. Both kind of objects typically will
25  * live inside the kernel with a refcnt of 2, one for its creation and one for
26  * the reference a group and a mark hold to each other.
27  * If you are holding the appropriate locks, you can take a reference and the
28  * object itself is guaranteed to survive until the reference is dropped.
29  *
30  * LOCKING:
31  * There are 3 locks involved with fsnotify inode marks and they MUST be taken
32  * in order as follows:
33  *
34  * group->mark_mutex
35  * mark->lock
36  * inode->i_lock
37  *
38  * group->mark_mutex protects the marks_list anchored inside a given group and
39  * each mark is hooked via the g_list.  It also protects the groups private
40  * data (i.e group limits).
41 
42  * mark->lock protects the marks attributes like its masks and flags.
43  * Furthermore it protects the access to a reference of the group that the mark
44  * is assigned to as well as the access to a reference of the inode/vfsmount
45  * that is being watched by the mark.
46  *
47  * inode->i_lock protects the i_fsnotify_marks list anchored inside a
48  * given inode and each mark is hooked via the i_list. (and sorta the
49  * free_i_list)
50  *
51  *
52  * LIFETIME:
53  * Inode marks survive between when they are added to an inode and when their
54  * refcnt==0.
55  *
56  * The inode mark can be cleared for a number of different reasons including:
57  * - The inode is unlinked for the last time.  (fsnotify_inode_remove)
58  * - The inode is being evicted from cache. (fsnotify_inode_delete)
59  * - The fs the inode is on is unmounted.  (fsnotify_inode_delete/fsnotify_unmount_inodes)
60  * - Something explicitly requests that it be removed.  (fsnotify_destroy_mark)
61  * - The fsnotify_group associated with the mark is going away and all such marks
62  *   need to be cleaned up. (fsnotify_clear_marks_by_group)
63  *
64  * Worst case we are given an inode and need to clean up all the marks on that
65  * inode.  We take i_lock and walk the i_fsnotify_marks safely.  For each
66  * mark on the list we take a reference (so the mark can't disappear under us).
67  * We remove that mark form the inode's list of marks and we add this mark to a
68  * private list anchored on the stack using i_free_list; we walk i_free_list
69  * and before we destroy the mark we make sure that we dont race with a
70  * concurrent destroy_group by getting a ref to the marks group and taking the
71  * groups mutex.
72 
73  * Very similarly for freeing by group, except we use free_g_list.
74  *
75  * This has the very interesting property of being able to run concurrently with
76  * any (or all) other directions.
77  */
78 
79 #include <linux/fs.h>
80 #include <linux/init.h>
81 #include <linux/kernel.h>
82 #include <linux/kthread.h>
83 #include <linux/module.h>
84 #include <linux/mutex.h>
85 #include <linux/slab.h>
86 #include <linux/spinlock.h>
87 #include <linux/srcu.h>
88 
89 #include <linux/atomic.h>
90 
91 #include <linux/fsnotify_backend.h>
92 #include "fsnotify.h"
93 
94 struct srcu_struct fsnotify_mark_srcu;
95 static DEFINE_SPINLOCK(destroy_lock);
96 static LIST_HEAD(destroy_list);
97 static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
98 
99 void fsnotify_get_mark(struct fsnotify_mark *mark)
100 {
101 	atomic_inc(&mark->refcnt);
102 }
103 
104 void fsnotify_put_mark(struct fsnotify_mark *mark)
105 {
106 	if (atomic_dec_and_test(&mark->refcnt)) {
107 		if (mark->group)
108 			fsnotify_put_group(mark->group);
109 		mark->free_mark(mark);
110 	}
111 }
112 
113 /* Calculate mask of events for a list of marks */
114 u32 fsnotify_recalc_mask(struct hlist_head *head)
115 {
116 	u32 new_mask = 0;
117 	struct fsnotify_mark *mark;
118 
119 	hlist_for_each_entry(mark, head, obj_list)
120 		new_mask |= mark->mask;
121 	return new_mask;
122 }
123 
124 /*
125  * Remove mark from inode / vfsmount list, group list, drop inode reference
126  * if we got one.
127  *
128  * Must be called with group->mark_mutex held.
129  */
130 void fsnotify_detach_mark(struct fsnotify_mark *mark)
131 {
132 	struct inode *inode = NULL;
133 	struct fsnotify_group *group = mark->group;
134 
135 	BUG_ON(!mutex_is_locked(&group->mark_mutex));
136 
137 	spin_lock(&mark->lock);
138 
139 	/* something else already called this function on this mark */
140 	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
141 		spin_unlock(&mark->lock);
142 		return;
143 	}
144 
145 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
146 
147 	if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
148 		inode = mark->inode;
149 		fsnotify_destroy_inode_mark(mark);
150 	} else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
151 		fsnotify_destroy_vfsmount_mark(mark);
152 	else
153 		BUG();
154 	/*
155 	 * Note that we didn't update flags telling whether inode cares about
156 	 * what's happening with children. We update these flags from
157 	 * __fsnotify_parent() lazily when next event happens on one of our
158 	 * children.
159 	 */
160 
161 	list_del_init(&mark->g_list);
162 
163 	spin_unlock(&mark->lock);
164 
165 	if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
166 		iput(inode);
167 
168 	atomic_dec(&group->num_marks);
169 }
170 
171 /*
172  * Free fsnotify mark. The freeing is actually happening from a kthread which
173  * first waits for srcu period end. Caller must have a reference to the mark
174  * or be protected by fsnotify_mark_srcu.
175  */
176 void fsnotify_free_mark(struct fsnotify_mark *mark)
177 {
178 	struct fsnotify_group *group = mark->group;
179 
180 	spin_lock(&mark->lock);
181 	/* something else already called this function on this mark */
182 	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
183 		spin_unlock(&mark->lock);
184 		return;
185 	}
186 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
187 	spin_unlock(&mark->lock);
188 
189 	spin_lock(&destroy_lock);
190 	list_add(&mark->g_list, &destroy_list);
191 	spin_unlock(&destroy_lock);
192 	wake_up(&destroy_waitq);
193 
194 	/*
195 	 * Some groups like to know that marks are being freed.  This is a
196 	 * callback to the group function to let it know that this mark
197 	 * is being freed.
198 	 */
199 	if (group->ops->freeing_mark)
200 		group->ops->freeing_mark(mark, group);
201 }
202 
203 void fsnotify_destroy_mark(struct fsnotify_mark *mark,
204 			   struct fsnotify_group *group)
205 {
206 	mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
207 	fsnotify_detach_mark(mark);
208 	mutex_unlock(&group->mark_mutex);
209 	fsnotify_free_mark(mark);
210 }
211 
212 void fsnotify_destroy_marks(struct hlist_head *head, spinlock_t *lock)
213 {
214 	struct fsnotify_mark *mark;
215 
216 	while (1) {
217 		/*
218 		 * We have to be careful since we can race with e.g.
219 		 * fsnotify_clear_marks_by_group() and once we drop 'lock',
220 		 * mark can get removed from the obj_list and destroyed. But
221 		 * we are holding mark reference so mark cannot be freed and
222 		 * calling fsnotify_destroy_mark() more than once is fine.
223 		 */
224 		spin_lock(lock);
225 		if (hlist_empty(head)) {
226 			spin_unlock(lock);
227 			break;
228 		}
229 		mark = hlist_entry(head->first, struct fsnotify_mark, obj_list);
230 		/*
231 		 * We don't update i_fsnotify_mask / mnt_fsnotify_mask here
232 		 * since inode / mount is going away anyway. So just remove
233 		 * mark from the list.
234 		 */
235 		hlist_del_init_rcu(&mark->obj_list);
236 		fsnotify_get_mark(mark);
237 		spin_unlock(lock);
238 		fsnotify_destroy_mark(mark, mark->group);
239 		fsnotify_put_mark(mark);
240 	}
241 }
242 
243 void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
244 {
245 	assert_spin_locked(&mark->lock);
246 
247 	mark->mask = mask;
248 
249 	if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
250 		fsnotify_set_inode_mark_mask_locked(mark, mask);
251 }
252 
253 void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
254 {
255 	assert_spin_locked(&mark->lock);
256 
257 	mark->ignored_mask = mask;
258 }
259 
260 /*
261  * Sorting function for lists of fsnotify marks.
262  *
263  * Fanotify supports different notification classes (reflected as priority of
264  * notification group). Events shall be passed to notification groups in
265  * decreasing priority order. To achieve this marks in notification lists for
266  * inodes and vfsmounts are sorted so that priorities of corresponding groups
267  * are descending.
268  *
269  * Furthermore correct handling of the ignore mask requires processing inode
270  * and vfsmount marks of each group together. Using the group address as
271  * further sort criterion provides a unique sorting order and thus we can
272  * merge inode and vfsmount lists of marks in linear time and find groups
273  * present in both lists.
274  *
275  * A return value of 1 signifies that b has priority over a.
276  * A return value of 0 signifies that the two marks have to be handled together.
277  * A return value of -1 signifies that a has priority over b.
278  */
279 int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
280 {
281 	if (a == b)
282 		return 0;
283 	if (!a)
284 		return 1;
285 	if (!b)
286 		return -1;
287 	if (a->priority < b->priority)
288 		return 1;
289 	if (a->priority > b->priority)
290 		return -1;
291 	if (a < b)
292 		return 1;
293 	return -1;
294 }
295 
296 /* Add mark into proper place in given list of marks */
297 int fsnotify_add_mark_list(struct hlist_head *head, struct fsnotify_mark *mark,
298 			   int allow_dups)
299 {
300 	struct fsnotify_mark *lmark, *last = NULL;
301 	int cmp;
302 
303 	/* is mark the first mark? */
304 	if (hlist_empty(head)) {
305 		hlist_add_head_rcu(&mark->obj_list, head);
306 		return 0;
307 	}
308 
309 	/* should mark be in the middle of the current list? */
310 	hlist_for_each_entry(lmark, head, obj_list) {
311 		last = lmark;
312 
313 		if ((lmark->group == mark->group) && !allow_dups)
314 			return -EEXIST;
315 
316 		cmp = fsnotify_compare_groups(lmark->group, mark->group);
317 		if (cmp >= 0) {
318 			hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
319 			return 0;
320 		}
321 	}
322 
323 	BUG_ON(last == NULL);
324 	/* mark should be the last entry.  last is the current last entry */
325 	hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
326 	return 0;
327 }
328 
329 /*
330  * Attach an initialized mark to a given group and fs object.
331  * These marks may be used for the fsnotify backend to determine which
332  * event types should be delivered to which group.
333  */
334 int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
335 			     struct fsnotify_group *group, struct inode *inode,
336 			     struct vfsmount *mnt, int allow_dups)
337 {
338 	int ret = 0;
339 
340 	BUG_ON(inode && mnt);
341 	BUG_ON(!inode && !mnt);
342 	BUG_ON(!mutex_is_locked(&group->mark_mutex));
343 
344 	/*
345 	 * LOCKING ORDER!!!!
346 	 * group->mark_mutex
347 	 * mark->lock
348 	 * inode->i_lock
349 	 */
350 	spin_lock(&mark->lock);
351 	mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
352 
353 	fsnotify_get_group(group);
354 	mark->group = group;
355 	list_add(&mark->g_list, &group->marks_list);
356 	atomic_inc(&group->num_marks);
357 	fsnotify_get_mark(mark); /* for i_list and g_list */
358 
359 	if (inode) {
360 		ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
361 		if (ret)
362 			goto err;
363 	} else if (mnt) {
364 		ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
365 		if (ret)
366 			goto err;
367 	} else {
368 		BUG();
369 	}
370 
371 	/* this will pin the object if appropriate */
372 	fsnotify_set_mark_mask_locked(mark, mark->mask);
373 	spin_unlock(&mark->lock);
374 
375 	if (inode)
376 		__fsnotify_update_child_dentry_flags(inode);
377 
378 	return ret;
379 err:
380 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
381 	list_del_init(&mark->g_list);
382 	fsnotify_put_group(group);
383 	mark->group = NULL;
384 	atomic_dec(&group->num_marks);
385 
386 	spin_unlock(&mark->lock);
387 
388 	spin_lock(&destroy_lock);
389 	list_add(&mark->g_list, &destroy_list);
390 	spin_unlock(&destroy_lock);
391 	wake_up(&destroy_waitq);
392 
393 	return ret;
394 }
395 
396 int fsnotify_add_mark(struct fsnotify_mark *mark, struct fsnotify_group *group,
397 		      struct inode *inode, struct vfsmount *mnt, int allow_dups)
398 {
399 	int ret;
400 	mutex_lock(&group->mark_mutex);
401 	ret = fsnotify_add_mark_locked(mark, group, inode, mnt, allow_dups);
402 	mutex_unlock(&group->mark_mutex);
403 	return ret;
404 }
405 
406 /*
407  * Given a list of marks, find the mark associated with given group. If found
408  * take a reference to that mark and return it, else return NULL.
409  */
410 struct fsnotify_mark *fsnotify_find_mark(struct hlist_head *head,
411 					 struct fsnotify_group *group)
412 {
413 	struct fsnotify_mark *mark;
414 
415 	hlist_for_each_entry(mark, head, obj_list) {
416 		if (mark->group == group) {
417 			fsnotify_get_mark(mark);
418 			return mark;
419 		}
420 	}
421 	return NULL;
422 }
423 
424 /*
425  * clear any marks in a group in which mark->flags & flags is true
426  */
427 void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
428 					 unsigned int flags)
429 {
430 	struct fsnotify_mark *lmark, *mark;
431 	LIST_HEAD(to_free);
432 
433 	/*
434 	 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
435 	 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
436 	 * to_free list so we have to use mark_mutex even when accessing that
437 	 * list. And freeing mark requires us to drop mark_mutex. So we can
438 	 * reliably free only the first mark in the list. That's why we first
439 	 * move marks to free to to_free list in one go and then free marks in
440 	 * to_free list one by one.
441 	 */
442 	mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
443 	list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
444 		if (mark->flags & flags)
445 			list_move(&mark->g_list, &to_free);
446 	}
447 	mutex_unlock(&group->mark_mutex);
448 
449 	while (1) {
450 		mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
451 		if (list_empty(&to_free)) {
452 			mutex_unlock(&group->mark_mutex);
453 			break;
454 		}
455 		mark = list_first_entry(&to_free, struct fsnotify_mark, g_list);
456 		fsnotify_get_mark(mark);
457 		fsnotify_detach_mark(mark);
458 		mutex_unlock(&group->mark_mutex);
459 		fsnotify_free_mark(mark);
460 		fsnotify_put_mark(mark);
461 	}
462 }
463 
464 /*
465  * Given a group, destroy all of the marks associated with that group.
466  */
467 void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
468 {
469 	fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
470 }
471 
472 void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
473 {
474 	assert_spin_locked(&old->lock);
475 	new->inode = old->inode;
476 	new->mnt = old->mnt;
477 	if (old->group)
478 		fsnotify_get_group(old->group);
479 	new->group = old->group;
480 	new->mask = old->mask;
481 	new->free_mark = old->free_mark;
482 }
483 
484 /*
485  * Nothing fancy, just initialize lists and locks and counters.
486  */
487 void fsnotify_init_mark(struct fsnotify_mark *mark,
488 			void (*free_mark)(struct fsnotify_mark *mark))
489 {
490 	memset(mark, 0, sizeof(*mark));
491 	spin_lock_init(&mark->lock);
492 	atomic_set(&mark->refcnt, 1);
493 	mark->free_mark = free_mark;
494 }
495 
496 static int fsnotify_mark_destroy(void *ignored)
497 {
498 	struct fsnotify_mark *mark, *next;
499 	struct list_head private_destroy_list;
500 
501 	for (;;) {
502 		spin_lock(&destroy_lock);
503 		/* exchange the list head */
504 		list_replace_init(&destroy_list, &private_destroy_list);
505 		spin_unlock(&destroy_lock);
506 
507 		synchronize_srcu(&fsnotify_mark_srcu);
508 
509 		list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
510 			list_del_init(&mark->g_list);
511 			fsnotify_put_mark(mark);
512 		}
513 
514 		wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
515 	}
516 
517 	return 0;
518 }
519 
520 static int __init fsnotify_mark_init(void)
521 {
522 	struct task_struct *thread;
523 
524 	thread = kthread_run(fsnotify_mark_destroy, NULL,
525 			     "fsnotify_mark");
526 	if (IS_ERR(thread))
527 		panic("unable to start fsnotify mark destruction thread.");
528 
529 	return 0;
530 }
531 device_initcall(fsnotify_mark_init);
532