Loading...
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 * mark->connector->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 * mark->connector->lock protects the list of marks anchored inside an
48 * inode / vfsmount and each mark is hooked via the i_list.
49 *
50 * A list of notification marks relating to inode / mnt is contained in
51 * fsnotify_mark_connector. That structure is alive as long as there are any
52 * marks in the list and is also protected by fsnotify_mark_srcu. A mark gets
53 * detached from fsnotify_mark_connector when last reference to the mark is
54 * dropped. Thus having mark reference is enough to protect mark->connector
55 * pointer and to make sure fsnotify_mark_connector cannot disappear. Also
56 * because we remove mark from g_list before dropping mark reference associated
57 * with that, any mark found through g_list is guaranteed to have
58 * mark->connector set until we drop group->mark_mutex.
59 *
60 * LIFETIME:
61 * Inode marks survive between when they are added to an inode and when their
62 * refcnt==0. Marks are also protected by fsnotify_mark_srcu.
63 *
64 * The inode mark can be cleared for a number of different reasons including:
65 * - The inode is unlinked for the last time. (fsnotify_inode_remove)
66 * - The inode is being evicted from cache. (fsnotify_inode_delete)
67 * - The fs the inode is on is unmounted. (fsnotify_inode_delete/fsnotify_unmount_inodes)
68 * - Something explicitly requests that it be removed. (fsnotify_destroy_mark)
69 * - The fsnotify_group associated with the mark is going away and all such marks
70 * need to be cleaned up. (fsnotify_clear_marks_by_group)
71 *
72 * This has the very interesting property of being able to run concurrently with
73 * any (or all) other directions.
74 */
75
76#include <linux/fs.h>
77#include <linux/init.h>
78#include <linux/kernel.h>
79#include <linux/kthread.h>
80#include <linux/module.h>
81#include <linux/mutex.h>
82#include <linux/slab.h>
83#include <linux/spinlock.h>
84#include <linux/srcu.h>
85
86#include <linux/atomic.h>
87
88#include <linux/fsnotify_backend.h>
89#include "fsnotify.h"
90
91#define FSNOTIFY_REAPER_DELAY (1) /* 1 jiffy */
92
93struct srcu_struct fsnotify_mark_srcu;
94struct kmem_cache *fsnotify_mark_connector_cachep;
95
96static DEFINE_SPINLOCK(destroy_lock);
97static LIST_HEAD(destroy_list);
98static struct fsnotify_mark_connector *connector_destroy_list;
99
100static void fsnotify_mark_destroy_workfn(struct work_struct *work);
101static DECLARE_DELAYED_WORK(reaper_work, fsnotify_mark_destroy_workfn);
102
103static void fsnotify_connector_destroy_workfn(struct work_struct *work);
104static DECLARE_WORK(connector_reaper_work, fsnotify_connector_destroy_workfn);
105
106void fsnotify_get_mark(struct fsnotify_mark *mark)
107{
108 WARN_ON_ONCE(!refcount_read(&mark->refcnt));
109 refcount_inc(&mark->refcnt);
110}
111
112static void __fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
113{
114 u32 new_mask = 0;
115 struct fsnotify_mark *mark;
116
117 assert_spin_locked(&conn->lock);
118 hlist_for_each_entry(mark, &conn->list, obj_list) {
119 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)
120 new_mask |= mark->mask;
121 }
122 if (conn->flags & FSNOTIFY_OBJ_TYPE_INODE)
123 conn->inode->i_fsnotify_mask = new_mask;
124 else if (conn->flags & FSNOTIFY_OBJ_TYPE_VFSMOUNT)
125 real_mount(conn->mnt)->mnt_fsnotify_mask = new_mask;
126}
127
128/*
129 * Calculate mask of events for a list of marks. The caller must make sure
130 * connector and connector->inode cannot disappear under us. Callers achieve
131 * this by holding a mark->lock or mark->group->mark_mutex for a mark on this
132 * list.
133 */
134void fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
135{
136 if (!conn)
137 return;
138
139 spin_lock(&conn->lock);
140 __fsnotify_recalc_mask(conn);
141 spin_unlock(&conn->lock);
142 if (conn->flags & FSNOTIFY_OBJ_TYPE_INODE)
143 __fsnotify_update_child_dentry_flags(conn->inode);
144}
145
146/* Free all connectors queued for freeing once SRCU period ends */
147static void fsnotify_connector_destroy_workfn(struct work_struct *work)
148{
149 struct fsnotify_mark_connector *conn, *free;
150
151 spin_lock(&destroy_lock);
152 conn = connector_destroy_list;
153 connector_destroy_list = NULL;
154 spin_unlock(&destroy_lock);
155
156 synchronize_srcu(&fsnotify_mark_srcu);
157 while (conn) {
158 free = conn;
159 conn = conn->destroy_next;
160 kmem_cache_free(fsnotify_mark_connector_cachep, free);
161 }
162}
163
164static struct inode *fsnotify_detach_connector_from_object(
165 struct fsnotify_mark_connector *conn)
166{
167 struct inode *inode = NULL;
168
169 if (conn->flags & FSNOTIFY_OBJ_TYPE_INODE) {
170 inode = conn->inode;
171 rcu_assign_pointer(inode->i_fsnotify_marks, NULL);
172 inode->i_fsnotify_mask = 0;
173 conn->inode = NULL;
174 conn->flags &= ~FSNOTIFY_OBJ_TYPE_INODE;
175 } else if (conn->flags & FSNOTIFY_OBJ_TYPE_VFSMOUNT) {
176 rcu_assign_pointer(real_mount(conn->mnt)->mnt_fsnotify_marks,
177 NULL);
178 real_mount(conn->mnt)->mnt_fsnotify_mask = 0;
179 conn->mnt = NULL;
180 conn->flags &= ~FSNOTIFY_OBJ_TYPE_VFSMOUNT;
181 }
182
183 return inode;
184}
185
186static void fsnotify_final_mark_destroy(struct fsnotify_mark *mark)
187{
188 struct fsnotify_group *group = mark->group;
189
190 if (WARN_ON_ONCE(!group))
191 return;
192 group->ops->free_mark(mark);
193 fsnotify_put_group(group);
194}
195
196void fsnotify_put_mark(struct fsnotify_mark *mark)
197{
198 struct fsnotify_mark_connector *conn;
199 struct inode *inode = NULL;
200 bool free_conn = false;
201
202 /* Catch marks that were actually never attached to object */
203 if (!mark->connector) {
204 if (refcount_dec_and_test(&mark->refcnt))
205 fsnotify_final_mark_destroy(mark);
206 return;
207 }
208
209 /*
210 * We have to be careful so that traversals of obj_list under lock can
211 * safely grab mark reference.
212 */
213 if (!refcount_dec_and_lock(&mark->refcnt, &mark->connector->lock))
214 return;
215
216 conn = mark->connector;
217 hlist_del_init_rcu(&mark->obj_list);
218 if (hlist_empty(&conn->list)) {
219 inode = fsnotify_detach_connector_from_object(conn);
220 free_conn = true;
221 } else {
222 __fsnotify_recalc_mask(conn);
223 }
224 mark->connector = NULL;
225 spin_unlock(&conn->lock);
226
227 iput(inode);
228
229 if (free_conn) {
230 spin_lock(&destroy_lock);
231 conn->destroy_next = connector_destroy_list;
232 connector_destroy_list = conn;
233 spin_unlock(&destroy_lock);
234 queue_work(system_unbound_wq, &connector_reaper_work);
235 }
236 /*
237 * Note that we didn't update flags telling whether inode cares about
238 * what's happening with children. We update these flags from
239 * __fsnotify_parent() lazily when next event happens on one of our
240 * children.
241 */
242 spin_lock(&destroy_lock);
243 list_add(&mark->g_list, &destroy_list);
244 spin_unlock(&destroy_lock);
245 queue_delayed_work(system_unbound_wq, &reaper_work,
246 FSNOTIFY_REAPER_DELAY);
247}
248
249/*
250 * Get mark reference when we found the mark via lockless traversal of object
251 * list. Mark can be already removed from the list by now and on its way to be
252 * destroyed once SRCU period ends.
253 *
254 * Also pin the group so it doesn't disappear under us.
255 */
256static bool fsnotify_get_mark_safe(struct fsnotify_mark *mark)
257{
258 if (!mark)
259 return true;
260
261 if (refcount_inc_not_zero(&mark->refcnt)) {
262 spin_lock(&mark->lock);
263 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) {
264 /* mark is attached, group is still alive then */
265 atomic_inc(&mark->group->user_waits);
266 spin_unlock(&mark->lock);
267 return true;
268 }
269 spin_unlock(&mark->lock);
270 fsnotify_put_mark(mark);
271 }
272 return false;
273}
274
275/*
276 * Puts marks and wakes up group destruction if necessary.
277 *
278 * Pairs with fsnotify_get_mark_safe()
279 */
280static void fsnotify_put_mark_wake(struct fsnotify_mark *mark)
281{
282 if (mark) {
283 struct fsnotify_group *group = mark->group;
284
285 fsnotify_put_mark(mark);
286 /*
287 * We abuse notification_waitq on group shutdown for waiting for
288 * all marks pinned when waiting for userspace.
289 */
290 if (atomic_dec_and_test(&group->user_waits) && group->shutdown)
291 wake_up(&group->notification_waitq);
292 }
293}
294
295bool fsnotify_prepare_user_wait(struct fsnotify_iter_info *iter_info)
296{
297 /* This can fail if mark is being removed */
298 if (!fsnotify_get_mark_safe(iter_info->inode_mark))
299 return false;
300 if (!fsnotify_get_mark_safe(iter_info->vfsmount_mark)) {
301 fsnotify_put_mark_wake(iter_info->inode_mark);
302 return false;
303 }
304
305 /*
306 * Now that both marks are pinned by refcount in the inode / vfsmount
307 * lists, we can drop SRCU lock, and safely resume the list iteration
308 * once userspace returns.
309 */
310 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
311
312 return true;
313}
314
315void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
316{
317 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
318 fsnotify_put_mark_wake(iter_info->inode_mark);
319 fsnotify_put_mark_wake(iter_info->vfsmount_mark);
320}
321
322/*
323 * Mark mark as detached, remove it from group list. Mark still stays in object
324 * list until its last reference is dropped. Note that we rely on mark being
325 * removed from group list before corresponding reference to it is dropped. In
326 * particular we rely on mark->connector being valid while we hold
327 * group->mark_mutex if we found the mark through g_list.
328 *
329 * Must be called with group->mark_mutex held. The caller must either hold
330 * reference to the mark or be protected by fsnotify_mark_srcu.
331 */
332void fsnotify_detach_mark(struct fsnotify_mark *mark)
333{
334 struct fsnotify_group *group = mark->group;
335
336 WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
337 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
338 refcount_read(&mark->refcnt) < 1 +
339 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
340
341 spin_lock(&mark->lock);
342 /* something else already called this function on this mark */
343 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
344 spin_unlock(&mark->lock);
345 return;
346 }
347 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
348 list_del_init(&mark->g_list);
349 spin_unlock(&mark->lock);
350
351 atomic_dec(&group->num_marks);
352
353 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
354 fsnotify_put_mark(mark);
355}
356
357/*
358 * Free fsnotify mark. The mark is actually only marked as being freed. The
359 * freeing is actually happening only once last reference to the mark is
360 * dropped from a workqueue which first waits for srcu period end.
361 *
362 * Caller must have a reference to the mark or be protected by
363 * fsnotify_mark_srcu.
364 */
365void fsnotify_free_mark(struct fsnotify_mark *mark)
366{
367 struct fsnotify_group *group = mark->group;
368
369 spin_lock(&mark->lock);
370 /* something else already called this function on this mark */
371 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
372 spin_unlock(&mark->lock);
373 return;
374 }
375 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
376 spin_unlock(&mark->lock);
377
378 /*
379 * Some groups like to know that marks are being freed. This is a
380 * callback to the group function to let it know that this mark
381 * is being freed.
382 */
383 if (group->ops->freeing_mark)
384 group->ops->freeing_mark(mark, group);
385}
386
387void fsnotify_destroy_mark(struct fsnotify_mark *mark,
388 struct fsnotify_group *group)
389{
390 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
391 fsnotify_detach_mark(mark);
392 mutex_unlock(&group->mark_mutex);
393 fsnotify_free_mark(mark);
394}
395
396/*
397 * Sorting function for lists of fsnotify marks.
398 *
399 * Fanotify supports different notification classes (reflected as priority of
400 * notification group). Events shall be passed to notification groups in
401 * decreasing priority order. To achieve this marks in notification lists for
402 * inodes and vfsmounts are sorted so that priorities of corresponding groups
403 * are descending.
404 *
405 * Furthermore correct handling of the ignore mask requires processing inode
406 * and vfsmount marks of each group together. Using the group address as
407 * further sort criterion provides a unique sorting order and thus we can
408 * merge inode and vfsmount lists of marks in linear time and find groups
409 * present in both lists.
410 *
411 * A return value of 1 signifies that b has priority over a.
412 * A return value of 0 signifies that the two marks have to be handled together.
413 * A return value of -1 signifies that a has priority over b.
414 */
415int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
416{
417 if (a == b)
418 return 0;
419 if (!a)
420 return 1;
421 if (!b)
422 return -1;
423 if (a->priority < b->priority)
424 return 1;
425 if (a->priority > b->priority)
426 return -1;
427 if (a < b)
428 return 1;
429 return -1;
430}
431
432static int fsnotify_attach_connector_to_object(
433 struct fsnotify_mark_connector __rcu **connp,
434 struct inode *inode,
435 struct vfsmount *mnt)
436{
437 struct fsnotify_mark_connector *conn;
438
439 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
440 if (!conn)
441 return -ENOMEM;
442 spin_lock_init(&conn->lock);
443 INIT_HLIST_HEAD(&conn->list);
444 if (inode) {
445 conn->flags = FSNOTIFY_OBJ_TYPE_INODE;
446 conn->inode = igrab(inode);
447 } else {
448 conn->flags = FSNOTIFY_OBJ_TYPE_VFSMOUNT;
449 conn->mnt = mnt;
450 }
451 /*
452 * cmpxchg() provides the barrier so that readers of *connp can see
453 * only initialized structure
454 */
455 if (cmpxchg(connp, NULL, conn)) {
456 /* Someone else created list structure for us */
457 if (inode)
458 iput(inode);
459 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
460 }
461
462 return 0;
463}
464
465/*
466 * Get mark connector, make sure it is alive and return with its lock held.
467 * This is for users that get connector pointer from inode or mount. Users that
468 * hold reference to a mark on the list may directly lock connector->lock as
469 * they are sure list cannot go away under them.
470 */
471static struct fsnotify_mark_connector *fsnotify_grab_connector(
472 struct fsnotify_mark_connector __rcu **connp)
473{
474 struct fsnotify_mark_connector *conn;
475 int idx;
476
477 idx = srcu_read_lock(&fsnotify_mark_srcu);
478 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
479 if (!conn)
480 goto out;
481 spin_lock(&conn->lock);
482 if (!(conn->flags & (FSNOTIFY_OBJ_TYPE_INODE |
483 FSNOTIFY_OBJ_TYPE_VFSMOUNT))) {
484 spin_unlock(&conn->lock);
485 srcu_read_unlock(&fsnotify_mark_srcu, idx);
486 return NULL;
487 }
488out:
489 srcu_read_unlock(&fsnotify_mark_srcu, idx);
490 return conn;
491}
492
493/*
494 * Add mark into proper place in given list of marks. These marks may be used
495 * for the fsnotify backend to determine which event types should be delivered
496 * to which group and for which inodes. These marks are ordered according to
497 * priority, highest number first, and then by the group's location in memory.
498 */
499static int fsnotify_add_mark_list(struct fsnotify_mark *mark,
500 struct inode *inode, struct vfsmount *mnt,
501 int allow_dups)
502{
503 struct fsnotify_mark *lmark, *last = NULL;
504 struct fsnotify_mark_connector *conn;
505 struct fsnotify_mark_connector __rcu **connp;
506 int cmp;
507 int err = 0;
508
509 if (WARN_ON(!inode && !mnt))
510 return -EINVAL;
511 if (inode)
512 connp = &inode->i_fsnotify_marks;
513 else
514 connp = &real_mount(mnt)->mnt_fsnotify_marks;
515restart:
516 spin_lock(&mark->lock);
517 conn = fsnotify_grab_connector(connp);
518 if (!conn) {
519 spin_unlock(&mark->lock);
520 err = fsnotify_attach_connector_to_object(connp, inode, mnt);
521 if (err)
522 return err;
523 goto restart;
524 }
525
526 /* is mark the first mark? */
527 if (hlist_empty(&conn->list)) {
528 hlist_add_head_rcu(&mark->obj_list, &conn->list);
529 goto added;
530 }
531
532 /* should mark be in the middle of the current list? */
533 hlist_for_each_entry(lmark, &conn->list, obj_list) {
534 last = lmark;
535
536 if ((lmark->group == mark->group) &&
537 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
538 !allow_dups) {
539 err = -EEXIST;
540 goto out_err;
541 }
542
543 cmp = fsnotify_compare_groups(lmark->group, mark->group);
544 if (cmp >= 0) {
545 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
546 goto added;
547 }
548 }
549
550 BUG_ON(last == NULL);
551 /* mark should be the last entry. last is the current last entry */
552 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
553added:
554 mark->connector = conn;
555out_err:
556 spin_unlock(&conn->lock);
557 spin_unlock(&mark->lock);
558 return err;
559}
560
561/*
562 * Attach an initialized mark to a given group and fs object.
563 * These marks may be used for the fsnotify backend to determine which
564 * event types should be delivered to which group.
565 */
566int fsnotify_add_mark_locked(struct fsnotify_mark *mark, struct inode *inode,
567 struct vfsmount *mnt, int allow_dups)
568{
569 struct fsnotify_group *group = mark->group;
570 int ret = 0;
571
572 BUG_ON(inode && mnt);
573 BUG_ON(!inode && !mnt);
574 BUG_ON(!mutex_is_locked(&group->mark_mutex));
575
576 /*
577 * LOCKING ORDER!!!!
578 * group->mark_mutex
579 * mark->lock
580 * mark->connector->lock
581 */
582 spin_lock(&mark->lock);
583 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
584
585 list_add(&mark->g_list, &group->marks_list);
586 atomic_inc(&group->num_marks);
587 fsnotify_get_mark(mark); /* for g_list */
588 spin_unlock(&mark->lock);
589
590 ret = fsnotify_add_mark_list(mark, inode, mnt, allow_dups);
591 if (ret)
592 goto err;
593
594 if (mark->mask)
595 fsnotify_recalc_mask(mark->connector);
596
597 return ret;
598err:
599 spin_lock(&mark->lock);
600 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
601 FSNOTIFY_MARK_FLAG_ATTACHED);
602 list_del_init(&mark->g_list);
603 spin_unlock(&mark->lock);
604 atomic_dec(&group->num_marks);
605
606 fsnotify_put_mark(mark);
607 return ret;
608}
609
610int fsnotify_add_mark(struct fsnotify_mark *mark, struct inode *inode,
611 struct vfsmount *mnt, int allow_dups)
612{
613 int ret;
614 struct fsnotify_group *group = mark->group;
615
616 mutex_lock(&group->mark_mutex);
617 ret = fsnotify_add_mark_locked(mark, inode, mnt, allow_dups);
618 mutex_unlock(&group->mark_mutex);
619 return ret;
620}
621
622/*
623 * Given a list of marks, find the mark associated with given group. If found
624 * take a reference to that mark and return it, else return NULL.
625 */
626struct fsnotify_mark *fsnotify_find_mark(
627 struct fsnotify_mark_connector __rcu **connp,
628 struct fsnotify_group *group)
629{
630 struct fsnotify_mark_connector *conn;
631 struct fsnotify_mark *mark;
632
633 conn = fsnotify_grab_connector(connp);
634 if (!conn)
635 return NULL;
636
637 hlist_for_each_entry(mark, &conn->list, obj_list) {
638 if (mark->group == group &&
639 (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
640 fsnotify_get_mark(mark);
641 spin_unlock(&conn->lock);
642 return mark;
643 }
644 }
645 spin_unlock(&conn->lock);
646 return NULL;
647}
648
649/* Clear any marks in a group with given type */
650void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
651 unsigned int type)
652{
653 struct fsnotify_mark *lmark, *mark;
654 LIST_HEAD(to_free);
655 struct list_head *head = &to_free;
656
657 /* Skip selection step if we want to clear all marks. */
658 if (type == FSNOTIFY_OBJ_ALL_TYPES) {
659 head = &group->marks_list;
660 goto clear;
661 }
662 /*
663 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
664 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
665 * to_free list so we have to use mark_mutex even when accessing that
666 * list. And freeing mark requires us to drop mark_mutex. So we can
667 * reliably free only the first mark in the list. That's why we first
668 * move marks to free to to_free list in one go and then free marks in
669 * to_free list one by one.
670 */
671 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
672 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
673 if (mark->connector->flags & type)
674 list_move(&mark->g_list, &to_free);
675 }
676 mutex_unlock(&group->mark_mutex);
677
678clear:
679 while (1) {
680 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
681 if (list_empty(head)) {
682 mutex_unlock(&group->mark_mutex);
683 break;
684 }
685 mark = list_first_entry(head, struct fsnotify_mark, g_list);
686 fsnotify_get_mark(mark);
687 fsnotify_detach_mark(mark);
688 mutex_unlock(&group->mark_mutex);
689 fsnotify_free_mark(mark);
690 fsnotify_put_mark(mark);
691 }
692}
693
694/* Destroy all marks attached to inode / vfsmount */
695void fsnotify_destroy_marks(struct fsnotify_mark_connector __rcu **connp)
696{
697 struct fsnotify_mark_connector *conn;
698 struct fsnotify_mark *mark, *old_mark = NULL;
699 struct inode *inode;
700
701 conn = fsnotify_grab_connector(connp);
702 if (!conn)
703 return;
704 /*
705 * We have to be careful since we can race with e.g.
706 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
707 * list can get modified. However we are holding mark reference and
708 * thus our mark cannot be removed from obj_list so we can continue
709 * iteration after regaining conn->lock.
710 */
711 hlist_for_each_entry(mark, &conn->list, obj_list) {
712 fsnotify_get_mark(mark);
713 spin_unlock(&conn->lock);
714 if (old_mark)
715 fsnotify_put_mark(old_mark);
716 old_mark = mark;
717 fsnotify_destroy_mark(mark, mark->group);
718 spin_lock(&conn->lock);
719 }
720 /*
721 * Detach list from object now so that we don't pin inode until all
722 * mark references get dropped. It would lead to strange results such
723 * as delaying inode deletion or blocking unmount.
724 */
725 inode = fsnotify_detach_connector_from_object(conn);
726 spin_unlock(&conn->lock);
727 if (old_mark)
728 fsnotify_put_mark(old_mark);
729 iput(inode);
730}
731
732/*
733 * Nothing fancy, just initialize lists and locks and counters.
734 */
735void fsnotify_init_mark(struct fsnotify_mark *mark,
736 struct fsnotify_group *group)
737{
738 memset(mark, 0, sizeof(*mark));
739 spin_lock_init(&mark->lock);
740 refcount_set(&mark->refcnt, 1);
741 fsnotify_get_group(group);
742 mark->group = group;
743}
744
745/*
746 * Destroy all marks in destroy_list, waits for SRCU period to finish before
747 * actually freeing marks.
748 */
749static void fsnotify_mark_destroy_workfn(struct work_struct *work)
750{
751 struct fsnotify_mark *mark, *next;
752 struct list_head private_destroy_list;
753
754 spin_lock(&destroy_lock);
755 /* exchange the list head */
756 list_replace_init(&destroy_list, &private_destroy_list);
757 spin_unlock(&destroy_lock);
758
759 synchronize_srcu(&fsnotify_mark_srcu);
760
761 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
762 list_del_init(&mark->g_list);
763 fsnotify_final_mark_destroy(mark);
764 }
765}
766
767/* Wait for all marks queued for destruction to be actually destroyed */
768void fsnotify_wait_marks_destroyed(void)
769{
770 flush_delayed_work(&reaper_work);
771}
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
100struct srcu_struct fsnotify_mark_srcu;
101static DEFINE_SPINLOCK(destroy_lock);
102static LIST_HEAD(destroy_list);
103static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
104
105void fsnotify_get_mark(struct fsnotify_mark *mark)
106{
107 atomic_inc(&mark->refcnt);
108}
109
110void 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 */
121void 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 /* 1 from caller and 1 for being on i_list/g_list */
139 BUG_ON(atomic_read(&mark->refcnt) < 2);
140
141 spin_lock(&group->mark_lock);
142
143 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
144 inode = mark->i.inode;
145 fsnotify_destroy_inode_mark(mark);
146 } else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
147 fsnotify_destroy_vfsmount_mark(mark);
148 else
149 BUG();
150
151 list_del_init(&mark->g_list);
152
153 spin_unlock(&group->mark_lock);
154 spin_unlock(&mark->lock);
155
156 spin_lock(&destroy_lock);
157 list_add(&mark->destroy_list, &destroy_list);
158 spin_unlock(&destroy_lock);
159 wake_up(&destroy_waitq);
160
161 /*
162 * Some groups like to know that marks are being freed. This is a
163 * callback to the group function to let it know that this mark
164 * is being freed.
165 */
166 if (group->ops->freeing_mark)
167 group->ops->freeing_mark(mark, group);
168
169 /*
170 * __fsnotify_update_child_dentry_flags(inode);
171 *
172 * I really want to call that, but we can't, we have no idea if the inode
173 * still exists the second we drop the mark->lock.
174 *
175 * The next time an event arrive to this inode from one of it's children
176 * __fsnotify_parent will see that the inode doesn't care about it's
177 * children and will update all of these flags then. So really this
178 * is just a lazy update (and could be a perf win...)
179 */
180
181 if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
182 iput(inode);
183
184 /*
185 * it's possible that this group tried to destroy itself, but this
186 * this mark was simultaneously being freed by inode. If that's the
187 * case, we finish freeing the group here.
188 */
189 if (unlikely(atomic_dec_and_test(&group->num_marks)))
190 fsnotify_final_destroy_group(group);
191}
192
193void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
194{
195 assert_spin_locked(&mark->lock);
196
197 mark->mask = mask;
198
199 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
200 fsnotify_set_inode_mark_mask_locked(mark, mask);
201}
202
203void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
204{
205 assert_spin_locked(&mark->lock);
206
207 mark->ignored_mask = mask;
208}
209
210/*
211 * Attach an initialized mark to a given group and fs object.
212 * These marks may be used for the fsnotify backend to determine which
213 * event types should be delivered to which group.
214 */
215int fsnotify_add_mark(struct fsnotify_mark *mark,
216 struct fsnotify_group *group, struct inode *inode,
217 struct vfsmount *mnt, int allow_dups)
218{
219 int ret = 0;
220
221 BUG_ON(inode && mnt);
222 BUG_ON(!inode && !mnt);
223
224 /*
225 * LOCKING ORDER!!!!
226 * mark->lock
227 * group->mark_lock
228 * inode->i_lock
229 */
230 spin_lock(&mark->lock);
231 spin_lock(&group->mark_lock);
232
233 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE;
234
235 mark->group = group;
236 list_add(&mark->g_list, &group->marks_list);
237 atomic_inc(&group->num_marks);
238 fsnotify_get_mark(mark); /* for i_list and g_list */
239
240 if (inode) {
241 ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
242 if (ret)
243 goto err;
244 } else if (mnt) {
245 ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
246 if (ret)
247 goto err;
248 } else {
249 BUG();
250 }
251
252 spin_unlock(&group->mark_lock);
253
254 /* this will pin the object if appropriate */
255 fsnotify_set_mark_mask_locked(mark, mark->mask);
256
257 spin_unlock(&mark->lock);
258
259 if (inode)
260 __fsnotify_update_child_dentry_flags(inode);
261
262 return ret;
263err:
264 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
265 list_del_init(&mark->g_list);
266 mark->group = NULL;
267 atomic_dec(&group->num_marks);
268
269 spin_unlock(&group->mark_lock);
270 spin_unlock(&mark->lock);
271
272 spin_lock(&destroy_lock);
273 list_add(&mark->destroy_list, &destroy_list);
274 spin_unlock(&destroy_lock);
275 wake_up(&destroy_waitq);
276
277 return ret;
278}
279
280/*
281 * clear any marks in a group in which mark->flags & flags is true
282 */
283void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
284 unsigned int flags)
285{
286 struct fsnotify_mark *lmark, *mark;
287 LIST_HEAD(free_list);
288
289 spin_lock(&group->mark_lock);
290 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
291 if (mark->flags & flags) {
292 list_add(&mark->free_g_list, &free_list);
293 list_del_init(&mark->g_list);
294 fsnotify_get_mark(mark);
295 }
296 }
297 spin_unlock(&group->mark_lock);
298
299 list_for_each_entry_safe(mark, lmark, &free_list, free_g_list) {
300 fsnotify_destroy_mark(mark);
301 fsnotify_put_mark(mark);
302 }
303}
304
305/*
306 * Given a group, destroy all of the marks associated with that group.
307 */
308void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
309{
310 fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
311}
312
313void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
314{
315 assert_spin_locked(&old->lock);
316 new->i.inode = old->i.inode;
317 new->m.mnt = old->m.mnt;
318 new->group = old->group;
319 new->mask = old->mask;
320 new->free_mark = old->free_mark;
321}
322
323/*
324 * Nothing fancy, just initialize lists and locks and counters.
325 */
326void fsnotify_init_mark(struct fsnotify_mark *mark,
327 void (*free_mark)(struct fsnotify_mark *mark))
328{
329 memset(mark, 0, sizeof(*mark));
330 spin_lock_init(&mark->lock);
331 atomic_set(&mark->refcnt, 1);
332 mark->free_mark = free_mark;
333}
334
335static int fsnotify_mark_destroy(void *ignored)
336{
337 struct fsnotify_mark *mark, *next;
338 LIST_HEAD(private_destroy_list);
339
340 for (;;) {
341 spin_lock(&destroy_lock);
342 /* exchange the list head */
343 list_replace_init(&destroy_list, &private_destroy_list);
344 spin_unlock(&destroy_lock);
345
346 synchronize_srcu(&fsnotify_mark_srcu);
347
348 list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) {
349 list_del_init(&mark->destroy_list);
350 fsnotify_put_mark(mark);
351 }
352
353 wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
354 }
355
356 return 0;
357}
358
359static int __init fsnotify_mark_init(void)
360{
361 struct task_struct *thread;
362
363 thread = kthread_run(fsnotify_mark_destroy, NULL,
364 "fsnotify_mark");
365 if (IS_ERR(thread))
366 panic("unable to start fsnotify mark destruction thread.");
367
368 return 0;
369}
370device_initcall(fsnotify_mark_init);