2 * Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
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)
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.
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.
20 * fsnotify inode mark locking/lifetime/and refcnting
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.
31 * There are 3 locks involved with fsnotify inode marks and they MUST be taken
32 * in order as follows:
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).
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.
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
53 * Inode marks survive between when they are added to an inode and when their
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)
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
73 * Very similarly for freeing by group, except we use free_g_list.
75 * This has the very interesting property of being able to run concurrently with
76 * any (or all) other directions.
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>
89 #include <linux/atomic.h>
91 #include <linux/fsnotify_backend.h>
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);
99 void fsnotify_get_mark(struct fsnotify_mark *mark)
101 atomic_inc(&mark->refcnt);
104 void fsnotify_put_mark(struct fsnotify_mark *mark)
106 if (atomic_dec_and_test(&mark->refcnt)) {
108 fsnotify_put_group(mark->group);
109 mark->free_mark(mark);
113 /* Calculate mask of events for a list of marks */
114 u32 fsnotify_recalc_mask(struct hlist_head *head)
117 struct fsnotify_mark *mark;
119 hlist_for_each_entry(mark, head, obj_list)
120 new_mask |= mark->mask;
125 * Any time a mark is getting freed we end up here.
126 * The caller had better be holding a reference to this mark so we don't actually
127 * do the final put under the mark->lock
129 void fsnotify_destroy_mark_locked(struct fsnotify_mark *mark,
130 struct fsnotify_group *group)
132 struct inode *inode = NULL;
134 BUG_ON(!mutex_is_locked(&group->mark_mutex));
136 spin_lock(&mark->lock);
138 /* something else already called this function on this mark */
139 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
140 spin_unlock(&mark->lock);
144 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
146 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
148 fsnotify_destroy_inode_mark(mark);
149 } else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
150 fsnotify_destroy_vfsmount_mark(mark);
154 list_del_init(&mark->g_list);
156 spin_unlock(&mark->lock);
158 if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
160 /* release lock temporarily */
161 mutex_unlock(&group->mark_mutex);
163 spin_lock(&destroy_lock);
164 list_add(&mark->g_list, &destroy_list);
165 spin_unlock(&destroy_lock);
166 wake_up(&destroy_waitq);
168 * We don't necessarily have a ref on mark from caller so the above destroy
169 * may have actually freed it, unless this group provides a 'freeing_mark'
170 * function which must be holding a reference.
174 * Some groups like to know that marks are being freed. This is a
175 * callback to the group function to let it know that this mark
178 if (group->ops->freeing_mark)
179 group->ops->freeing_mark(mark, group);
182 * __fsnotify_update_child_dentry_flags(inode);
184 * I really want to call that, but we can't, we have no idea if the inode
185 * still exists the second we drop the mark->lock.
187 * The next time an event arrive to this inode from one of it's children
188 * __fsnotify_parent will see that the inode doesn't care about it's
189 * children and will update all of these flags then. So really this
190 * is just a lazy update (and could be a perf win...)
193 atomic_dec(&group->num_marks);
195 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
198 void fsnotify_destroy_mark(struct fsnotify_mark *mark,
199 struct fsnotify_group *group)
201 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
202 fsnotify_destroy_mark_locked(mark, group);
203 mutex_unlock(&group->mark_mutex);
206 void fsnotify_destroy_marks(struct hlist_head *head, spinlock_t *lock)
208 struct fsnotify_mark *mark;
212 * We have to be careful since we can race with e.g.
213 * fsnotify_clear_marks_by_group() and once we drop 'lock',
214 * mark can get removed from the obj_list and destroyed. But
215 * we are holding mark reference so mark cannot be freed and
216 * calling fsnotify_destroy_mark() more than once is fine.
219 if (hlist_empty(head)) {
223 mark = hlist_entry(head->first, struct fsnotify_mark, obj_list);
225 * We don't update i_fsnotify_mask / mnt_fsnotify_mask here
226 * since inode / mount is going away anyway. So just remove
227 * mark from the list.
229 hlist_del_init_rcu(&mark->obj_list);
230 fsnotify_get_mark(mark);
232 fsnotify_destroy_mark(mark, mark->group);
233 fsnotify_put_mark(mark);
237 void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
239 assert_spin_locked(&mark->lock);
243 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
244 fsnotify_set_inode_mark_mask_locked(mark, mask);
247 void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
249 assert_spin_locked(&mark->lock);
251 mark->ignored_mask = mask;
255 * Sorting function for lists of fsnotify marks.
257 * Fanotify supports different notification classes (reflected as priority of
258 * notification group). Events shall be passed to notification groups in
259 * decreasing priority order. To achieve this marks in notification lists for
260 * inodes and vfsmounts are sorted so that priorities of corresponding groups
263 * Furthermore correct handling of the ignore mask requires processing inode
264 * and vfsmount marks of each group together. Using the group address as
265 * further sort criterion provides a unique sorting order and thus we can
266 * merge inode and vfsmount lists of marks in linear time and find groups
267 * present in both lists.
269 * A return value of 1 signifies that b has priority over a.
270 * A return value of 0 signifies that the two marks have to be handled together.
271 * A return value of -1 signifies that a has priority over b.
273 int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
281 if (a->priority < b->priority)
283 if (a->priority > b->priority)
290 /* Add mark into proper place in given list of marks */
291 int fsnotify_add_mark_list(struct hlist_head *head, struct fsnotify_mark *mark,
294 struct fsnotify_mark *lmark, *last = NULL;
297 /* is mark the first mark? */
298 if (hlist_empty(head)) {
299 hlist_add_head_rcu(&mark->obj_list, head);
303 /* should mark be in the middle of the current list? */
304 hlist_for_each_entry(lmark, head, obj_list) {
307 if ((lmark->group == mark->group) && !allow_dups)
310 cmp = fsnotify_compare_groups(lmark->group, mark->group);
312 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
317 BUG_ON(last == NULL);
318 /* mark should be the last entry. last is the current last entry */
319 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
324 * Attach an initialized mark to a given group and fs object.
325 * These marks may be used for the fsnotify backend to determine which
326 * event types should be delivered to which group.
328 int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
329 struct fsnotify_group *group, struct inode *inode,
330 struct vfsmount *mnt, int allow_dups)
334 BUG_ON(inode && mnt);
335 BUG_ON(!inode && !mnt);
336 BUG_ON(!mutex_is_locked(&group->mark_mutex));
344 spin_lock(&mark->lock);
345 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE;
347 fsnotify_get_group(group);
349 list_add(&mark->g_list, &group->marks_list);
350 atomic_inc(&group->num_marks);
351 fsnotify_get_mark(mark); /* for i_list and g_list */
354 ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
358 ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
365 /* this will pin the object if appropriate */
366 fsnotify_set_mark_mask_locked(mark, mark->mask);
367 spin_unlock(&mark->lock);
370 __fsnotify_update_child_dentry_flags(inode);
374 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
375 list_del_init(&mark->g_list);
376 fsnotify_put_group(group);
378 atomic_dec(&group->num_marks);
380 spin_unlock(&mark->lock);
382 spin_lock(&destroy_lock);
383 list_add(&mark->g_list, &destroy_list);
384 spin_unlock(&destroy_lock);
385 wake_up(&destroy_waitq);
390 int fsnotify_add_mark(struct fsnotify_mark *mark, struct fsnotify_group *group,
391 struct inode *inode, struct vfsmount *mnt, int allow_dups)
394 mutex_lock(&group->mark_mutex);
395 ret = fsnotify_add_mark_locked(mark, group, inode, mnt, allow_dups);
396 mutex_unlock(&group->mark_mutex);
401 * Given a list of marks, find the mark associated with given group. If found
402 * take a reference to that mark and return it, else return NULL.
404 struct fsnotify_mark *fsnotify_find_mark(struct hlist_head *head,
405 struct fsnotify_group *group)
407 struct fsnotify_mark *mark;
409 hlist_for_each_entry(mark, head, obj_list) {
410 if (mark->group == group) {
411 fsnotify_get_mark(mark);
419 * clear any marks in a group in which mark->flags & flags is true
421 void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
424 struct fsnotify_mark *lmark, *mark;
428 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
429 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
430 * to_free list so we have to use mark_mutex even when accessing that
431 * list. And freeing mark requires us to drop mark_mutex. So we can
432 * reliably free only the first mark in the list. That's why we first
433 * move marks to free to to_free list in one go and then free marks in
434 * to_free list one by one.
436 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
437 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
438 if (mark->flags & flags)
439 list_move(&mark->g_list, &to_free);
441 mutex_unlock(&group->mark_mutex);
444 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
445 if (list_empty(&to_free)) {
446 mutex_unlock(&group->mark_mutex);
449 mark = list_first_entry(&to_free, struct fsnotify_mark, g_list);
450 fsnotify_get_mark(mark);
451 fsnotify_destroy_mark_locked(mark, group);
452 mutex_unlock(&group->mark_mutex);
453 fsnotify_put_mark(mark);
458 * Given a group, destroy all of the marks associated with that group.
460 void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
462 fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
465 void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
467 assert_spin_locked(&old->lock);
468 new->inode = old->inode;
471 fsnotify_get_group(old->group);
472 new->group = old->group;
473 new->mask = old->mask;
474 new->free_mark = old->free_mark;
478 * Nothing fancy, just initialize lists and locks and counters.
480 void fsnotify_init_mark(struct fsnotify_mark *mark,
481 void (*free_mark)(struct fsnotify_mark *mark))
483 memset(mark, 0, sizeof(*mark));
484 spin_lock_init(&mark->lock);
485 atomic_set(&mark->refcnt, 1);
486 mark->free_mark = free_mark;
489 static int fsnotify_mark_destroy(void *ignored)
491 struct fsnotify_mark *mark, *next;
492 struct list_head private_destroy_list;
495 spin_lock(&destroy_lock);
496 /* exchange the list head */
497 list_replace_init(&destroy_list, &private_destroy_list);
498 spin_unlock(&destroy_lock);
500 synchronize_srcu(&fsnotify_mark_srcu);
502 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
503 list_del_init(&mark->g_list);
504 fsnotify_put_mark(mark);
507 wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
513 static int __init fsnotify_mark_init(void)
515 struct task_struct *thread;
517 thread = kthread_run(fsnotify_mark_destroy, NULL,
520 panic("unable to start fsnotify mark destruction thread.");
524 device_initcall(fsnotify_mark_init);