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// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
4 */
5
6/*
7 * fsnotify inode mark locking/lifetime/and refcnting
8 *
9 * REFCNT:
10 * The group->recnt and mark->refcnt tell how many "things" in the kernel
11 * currently are referencing the objects. Both kind of objects typically will
12 * live inside the kernel with a refcnt of 2, one for its creation and one for
13 * the reference a group and a mark hold to each other.
14 * If you are holding the appropriate locks, you can take a reference and the
15 * object itself is guaranteed to survive until the reference is dropped.
16 *
17 * LOCKING:
18 * There are 3 locks involved with fsnotify inode marks and they MUST be taken
19 * in order as follows:
20 *
21 * group->mark_mutex
22 * mark->lock
23 * mark->connector->lock
24 *
25 * group->mark_mutex protects the marks_list anchored inside a given group and
26 * each mark is hooked via the g_list. It also protects the groups private
27 * data (i.e group limits).
28
29 * mark->lock protects the marks attributes like its masks and flags.
30 * Furthermore it protects the access to a reference of the group that the mark
31 * is assigned to as well as the access to a reference of the inode/vfsmount
32 * that is being watched by the mark.
33 *
34 * mark->connector->lock protects the list of marks anchored inside an
35 * inode / vfsmount and each mark is hooked via the i_list.
36 *
37 * A list of notification marks relating to inode / mnt is contained in
38 * fsnotify_mark_connector. That structure is alive as long as there are any
39 * marks in the list and is also protected by fsnotify_mark_srcu. A mark gets
40 * detached from fsnotify_mark_connector when last reference to the mark is
41 * dropped. Thus having mark reference is enough to protect mark->connector
42 * pointer and to make sure fsnotify_mark_connector cannot disappear. Also
43 * because we remove mark from g_list before dropping mark reference associated
44 * with that, any mark found through g_list is guaranteed to have
45 * mark->connector set until we drop group->mark_mutex.
46 *
47 * LIFETIME:
48 * Inode marks survive between when they are added to an inode and when their
49 * refcnt==0. Marks are also protected by fsnotify_mark_srcu.
50 *
51 * The inode mark can be cleared for a number of different reasons including:
52 * - The inode is unlinked for the last time. (fsnotify_inode_remove)
53 * - The inode is being evicted from cache. (fsnotify_inode_delete)
54 * - The fs the inode is on is unmounted. (fsnotify_inode_delete/fsnotify_unmount_inodes)
55 * - Something explicitly requests that it be removed. (fsnotify_destroy_mark)
56 * - The fsnotify_group associated with the mark is going away and all such marks
57 * need to be cleaned up. (fsnotify_clear_marks_by_group)
58 *
59 * This has the very interesting property of being able to run concurrently with
60 * any (or all) other directions.
61 */
62
63#include <linux/fs.h>
64#include <linux/init.h>
65#include <linux/kernel.h>
66#include <linux/kthread.h>
67#include <linux/module.h>
68#include <linux/mutex.h>
69#include <linux/slab.h>
70#include <linux/spinlock.h>
71#include <linux/srcu.h>
72#include <linux/ratelimit.h>
73
74#include <linux/atomic.h>
75
76#include <linux/fsnotify_backend.h>
77#include "fsnotify.h"
78
79#define FSNOTIFY_REAPER_DELAY (1) /* 1 jiffy */
80
81struct srcu_struct fsnotify_mark_srcu;
82struct kmem_cache *fsnotify_mark_connector_cachep;
83
84static DEFINE_SPINLOCK(destroy_lock);
85static LIST_HEAD(destroy_list);
86static struct fsnotify_mark_connector *connector_destroy_list;
87
88static void fsnotify_mark_destroy_workfn(struct work_struct *work);
89static DECLARE_DELAYED_WORK(reaper_work, fsnotify_mark_destroy_workfn);
90
91static void fsnotify_connector_destroy_workfn(struct work_struct *work);
92static DECLARE_WORK(connector_reaper_work, fsnotify_connector_destroy_workfn);
93
94void fsnotify_get_mark(struct fsnotify_mark *mark)
95{
96 WARN_ON_ONCE(!refcount_read(&mark->refcnt));
97 refcount_inc(&mark->refcnt);
98}
99
100static fsnotify_connp_t *fsnotify_object_connp(void *obj,
101 enum fsnotify_obj_type obj_type)
102{
103 switch (obj_type) {
104 case FSNOTIFY_OBJ_TYPE_INODE:
105 return &((struct inode *)obj)->i_fsnotify_marks;
106 case FSNOTIFY_OBJ_TYPE_VFSMOUNT:
107 return &real_mount(obj)->mnt_fsnotify_marks;
108 case FSNOTIFY_OBJ_TYPE_SB:
109 return fsnotify_sb_marks(obj);
110 default:
111 return NULL;
112 }
113}
114
115static __u32 *fsnotify_conn_mask_p(struct fsnotify_mark_connector *conn)
116{
117 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
118 return &fsnotify_conn_inode(conn)->i_fsnotify_mask;
119 else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT)
120 return &fsnotify_conn_mount(conn)->mnt_fsnotify_mask;
121 else if (conn->type == FSNOTIFY_OBJ_TYPE_SB)
122 return &fsnotify_conn_sb(conn)->s_fsnotify_mask;
123 return NULL;
124}
125
126__u32 fsnotify_conn_mask(struct fsnotify_mark_connector *conn)
127{
128 if (WARN_ON(!fsnotify_valid_obj_type(conn->type)))
129 return 0;
130
131 return READ_ONCE(*fsnotify_conn_mask_p(conn));
132}
133
134static void fsnotify_get_sb_watched_objects(struct super_block *sb)
135{
136 atomic_long_inc(fsnotify_sb_watched_objects(sb));
137}
138
139static void fsnotify_put_sb_watched_objects(struct super_block *sb)
140{
141 atomic_long_t *watched_objects = fsnotify_sb_watched_objects(sb);
142
143 /* the superblock can go away after this decrement */
144 if (atomic_long_dec_and_test(watched_objects))
145 wake_up_var(watched_objects);
146}
147
148static void fsnotify_get_inode_ref(struct inode *inode)
149{
150 ihold(inode);
151 fsnotify_get_sb_watched_objects(inode->i_sb);
152}
153
154static void fsnotify_put_inode_ref(struct inode *inode)
155{
156 /* read ->i_sb before the inode can go away */
157 struct super_block *sb = inode->i_sb;
158
159 iput(inode);
160 fsnotify_put_sb_watched_objects(sb);
161}
162
163/*
164 * Grab or drop watched objects reference depending on whether the connector
165 * is attached and has any marks attached.
166 */
167static void fsnotify_update_sb_watchers(struct super_block *sb,
168 struct fsnotify_mark_connector *conn)
169{
170 struct fsnotify_sb_info *sbinfo = fsnotify_sb_info(sb);
171 bool is_watched = conn->flags & FSNOTIFY_CONN_FLAG_IS_WATCHED;
172 struct fsnotify_mark *first_mark = NULL;
173 unsigned int highest_prio = 0;
174
175 if (conn->obj)
176 first_mark = hlist_entry_safe(conn->list.first,
177 struct fsnotify_mark, obj_list);
178 if (first_mark)
179 highest_prio = first_mark->group->priority;
180 if (WARN_ON(highest_prio >= __FSNOTIFY_PRIO_NUM))
181 highest_prio = 0;
182
183 /*
184 * If the highest priority of group watching this object is prio,
185 * then watched object has a reference on counters [0..prio].
186 * Update priority >= 1 watched objects counters.
187 */
188 for (unsigned int p = conn->prio + 1; p <= highest_prio; p++)
189 atomic_long_inc(&sbinfo->watched_objects[p]);
190 for (unsigned int p = conn->prio; p > highest_prio; p--)
191 atomic_long_dec(&sbinfo->watched_objects[p]);
192 conn->prio = highest_prio;
193
194 /* Update priority >= 0 (a.k.a total) watched objects counter */
195 BUILD_BUG_ON(FSNOTIFY_PRIO_NORMAL != 0);
196 if (first_mark && !is_watched) {
197 conn->flags |= FSNOTIFY_CONN_FLAG_IS_WATCHED;
198 fsnotify_get_sb_watched_objects(sb);
199 } else if (!first_mark && is_watched) {
200 conn->flags &= ~FSNOTIFY_CONN_FLAG_IS_WATCHED;
201 fsnotify_put_sb_watched_objects(sb);
202 }
203}
204
205/*
206 * Grab or drop inode reference for the connector if needed.
207 *
208 * When it's time to drop the reference, we only clear the HAS_IREF flag and
209 * return the inode object. fsnotify_drop_object() will be resonsible for doing
210 * iput() outside of spinlocks. This happens when last mark that wanted iref is
211 * detached.
212 */
213static struct inode *fsnotify_update_iref(struct fsnotify_mark_connector *conn,
214 bool want_iref)
215{
216 bool has_iref = conn->flags & FSNOTIFY_CONN_FLAG_HAS_IREF;
217 struct inode *inode = NULL;
218
219 if (conn->type != FSNOTIFY_OBJ_TYPE_INODE ||
220 want_iref == has_iref)
221 return NULL;
222
223 if (want_iref) {
224 /* Pin inode if any mark wants inode refcount held */
225 fsnotify_get_inode_ref(fsnotify_conn_inode(conn));
226 conn->flags |= FSNOTIFY_CONN_FLAG_HAS_IREF;
227 } else {
228 /* Unpin inode after detach of last mark that wanted iref */
229 inode = fsnotify_conn_inode(conn);
230 conn->flags &= ~FSNOTIFY_CONN_FLAG_HAS_IREF;
231 }
232
233 return inode;
234}
235
236static void *__fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
237{
238 u32 new_mask = 0;
239 bool want_iref = false;
240 struct fsnotify_mark *mark;
241
242 assert_spin_locked(&conn->lock);
243 /* We can get detached connector here when inode is getting unlinked. */
244 if (!fsnotify_valid_obj_type(conn->type))
245 return NULL;
246 hlist_for_each_entry(mark, &conn->list, obj_list) {
247 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED))
248 continue;
249 new_mask |= fsnotify_calc_mask(mark);
250 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE &&
251 !(mark->flags & FSNOTIFY_MARK_FLAG_NO_IREF))
252 want_iref = true;
253 }
254 /*
255 * We use WRITE_ONCE() to prevent silly compiler optimizations from
256 * confusing readers not holding conn->lock with partial updates.
257 */
258 WRITE_ONCE(*fsnotify_conn_mask_p(conn), new_mask);
259
260 return fsnotify_update_iref(conn, want_iref);
261}
262
263static bool fsnotify_conn_watches_children(
264 struct fsnotify_mark_connector *conn)
265{
266 if (conn->type != FSNOTIFY_OBJ_TYPE_INODE)
267 return false;
268
269 return fsnotify_inode_watches_children(fsnotify_conn_inode(conn));
270}
271
272static void fsnotify_conn_set_children_dentry_flags(
273 struct fsnotify_mark_connector *conn)
274{
275 if (conn->type != FSNOTIFY_OBJ_TYPE_INODE)
276 return;
277
278 fsnotify_set_children_dentry_flags(fsnotify_conn_inode(conn));
279}
280
281/*
282 * Calculate mask of events for a list of marks. The caller must make sure
283 * connector and connector->obj cannot disappear under us. Callers achieve
284 * this by holding a mark->lock or mark->group->mark_mutex for a mark on this
285 * list.
286 */
287void fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
288{
289 bool update_children;
290
291 if (!conn)
292 return;
293
294 spin_lock(&conn->lock);
295 update_children = !fsnotify_conn_watches_children(conn);
296 __fsnotify_recalc_mask(conn);
297 update_children &= fsnotify_conn_watches_children(conn);
298 spin_unlock(&conn->lock);
299 /*
300 * Set children's PARENT_WATCHED flags only if parent started watching.
301 * When parent stops watching, we clear false positive PARENT_WATCHED
302 * flags lazily in __fsnotify_parent().
303 */
304 if (update_children)
305 fsnotify_conn_set_children_dentry_flags(conn);
306}
307
308/* Free all connectors queued for freeing once SRCU period ends */
309static void fsnotify_connector_destroy_workfn(struct work_struct *work)
310{
311 struct fsnotify_mark_connector *conn, *free;
312
313 spin_lock(&destroy_lock);
314 conn = connector_destroy_list;
315 connector_destroy_list = NULL;
316 spin_unlock(&destroy_lock);
317
318 synchronize_srcu(&fsnotify_mark_srcu);
319 while (conn) {
320 free = conn;
321 conn = conn->destroy_next;
322 kmem_cache_free(fsnotify_mark_connector_cachep, free);
323 }
324}
325
326static void *fsnotify_detach_connector_from_object(
327 struct fsnotify_mark_connector *conn,
328 unsigned int *type)
329{
330 fsnotify_connp_t *connp = fsnotify_object_connp(conn->obj, conn->type);
331 struct super_block *sb = fsnotify_connector_sb(conn);
332 struct inode *inode = NULL;
333
334 *type = conn->type;
335 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED)
336 return NULL;
337
338 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE) {
339 inode = fsnotify_conn_inode(conn);
340 inode->i_fsnotify_mask = 0;
341
342 /* Unpin inode when detaching from connector */
343 if (!(conn->flags & FSNOTIFY_CONN_FLAG_HAS_IREF))
344 inode = NULL;
345 } else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT) {
346 fsnotify_conn_mount(conn)->mnt_fsnotify_mask = 0;
347 } else if (conn->type == FSNOTIFY_OBJ_TYPE_SB) {
348 fsnotify_conn_sb(conn)->s_fsnotify_mask = 0;
349 }
350
351 rcu_assign_pointer(*connp, NULL);
352 conn->obj = NULL;
353 conn->type = FSNOTIFY_OBJ_TYPE_DETACHED;
354 fsnotify_update_sb_watchers(sb, conn);
355
356 return inode;
357}
358
359static void fsnotify_final_mark_destroy(struct fsnotify_mark *mark)
360{
361 struct fsnotify_group *group = mark->group;
362
363 if (WARN_ON_ONCE(!group))
364 return;
365 group->ops->free_mark(mark);
366 fsnotify_put_group(group);
367}
368
369/* Drop object reference originally held by a connector */
370static void fsnotify_drop_object(unsigned int type, void *objp)
371{
372 if (!objp)
373 return;
374 /* Currently only inode references are passed to be dropped */
375 if (WARN_ON_ONCE(type != FSNOTIFY_OBJ_TYPE_INODE))
376 return;
377 fsnotify_put_inode_ref(objp);
378}
379
380void fsnotify_put_mark(struct fsnotify_mark *mark)
381{
382 struct fsnotify_mark_connector *conn = READ_ONCE(mark->connector);
383 void *objp = NULL;
384 unsigned int type = FSNOTIFY_OBJ_TYPE_DETACHED;
385 bool free_conn = false;
386
387 /* Catch marks that were actually never attached to object */
388 if (!conn) {
389 if (refcount_dec_and_test(&mark->refcnt))
390 fsnotify_final_mark_destroy(mark);
391 return;
392 }
393
394 /*
395 * We have to be careful so that traversals of obj_list under lock can
396 * safely grab mark reference.
397 */
398 if (!refcount_dec_and_lock(&mark->refcnt, &conn->lock))
399 return;
400
401 hlist_del_init_rcu(&mark->obj_list);
402 if (hlist_empty(&conn->list)) {
403 objp = fsnotify_detach_connector_from_object(conn, &type);
404 free_conn = true;
405 } else {
406 struct super_block *sb = fsnotify_connector_sb(conn);
407
408 /* Update watched objects after detaching mark */
409 if (sb)
410 fsnotify_update_sb_watchers(sb, conn);
411 objp = __fsnotify_recalc_mask(conn);
412 type = conn->type;
413 }
414 WRITE_ONCE(mark->connector, NULL);
415 spin_unlock(&conn->lock);
416
417 fsnotify_drop_object(type, objp);
418
419 if (free_conn) {
420 spin_lock(&destroy_lock);
421 conn->destroy_next = connector_destroy_list;
422 connector_destroy_list = conn;
423 spin_unlock(&destroy_lock);
424 queue_work(system_unbound_wq, &connector_reaper_work);
425 }
426 /*
427 * Note that we didn't update flags telling whether inode cares about
428 * what's happening with children. We update these flags from
429 * __fsnotify_parent() lazily when next event happens on one of our
430 * children.
431 */
432 spin_lock(&destroy_lock);
433 list_add(&mark->g_list, &destroy_list);
434 spin_unlock(&destroy_lock);
435 queue_delayed_work(system_unbound_wq, &reaper_work,
436 FSNOTIFY_REAPER_DELAY);
437}
438EXPORT_SYMBOL_GPL(fsnotify_put_mark);
439
440/*
441 * Get mark reference when we found the mark via lockless traversal of object
442 * list. Mark can be already removed from the list by now and on its way to be
443 * destroyed once SRCU period ends.
444 *
445 * Also pin the group so it doesn't disappear under us.
446 */
447static bool fsnotify_get_mark_safe(struct fsnotify_mark *mark)
448{
449 if (!mark)
450 return true;
451
452 if (refcount_inc_not_zero(&mark->refcnt)) {
453 spin_lock(&mark->lock);
454 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) {
455 /* mark is attached, group is still alive then */
456 atomic_inc(&mark->group->user_waits);
457 spin_unlock(&mark->lock);
458 return true;
459 }
460 spin_unlock(&mark->lock);
461 fsnotify_put_mark(mark);
462 }
463 return false;
464}
465
466/*
467 * Puts marks and wakes up group destruction if necessary.
468 *
469 * Pairs with fsnotify_get_mark_safe()
470 */
471static void fsnotify_put_mark_wake(struct fsnotify_mark *mark)
472{
473 if (mark) {
474 struct fsnotify_group *group = mark->group;
475
476 fsnotify_put_mark(mark);
477 /*
478 * We abuse notification_waitq on group shutdown for waiting for
479 * all marks pinned when waiting for userspace.
480 */
481 if (atomic_dec_and_test(&group->user_waits) && group->shutdown)
482 wake_up(&group->notification_waitq);
483 }
484}
485
486bool fsnotify_prepare_user_wait(struct fsnotify_iter_info *iter_info)
487 __releases(&fsnotify_mark_srcu)
488{
489 int type;
490
491 fsnotify_foreach_iter_type(type) {
492 /* This can fail if mark is being removed */
493 if (!fsnotify_get_mark_safe(iter_info->marks[type])) {
494 __release(&fsnotify_mark_srcu);
495 goto fail;
496 }
497 }
498
499 /*
500 * Now that both marks are pinned by refcount in the inode / vfsmount
501 * lists, we can drop SRCU lock, and safely resume the list iteration
502 * once userspace returns.
503 */
504 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
505
506 return true;
507
508fail:
509 for (type--; type >= 0; type--)
510 fsnotify_put_mark_wake(iter_info->marks[type]);
511 return false;
512}
513
514void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
515 __acquires(&fsnotify_mark_srcu)
516{
517 int type;
518
519 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
520 fsnotify_foreach_iter_type(type)
521 fsnotify_put_mark_wake(iter_info->marks[type]);
522}
523
524/*
525 * Mark mark as detached, remove it from group list. Mark still stays in object
526 * list until its last reference is dropped. Note that we rely on mark being
527 * removed from group list before corresponding reference to it is dropped. In
528 * particular we rely on mark->connector being valid while we hold
529 * group->mark_mutex if we found the mark through g_list.
530 *
531 * Must be called with group->mark_mutex held. The caller must either hold
532 * reference to the mark or be protected by fsnotify_mark_srcu.
533 */
534void fsnotify_detach_mark(struct fsnotify_mark *mark)
535{
536 fsnotify_group_assert_locked(mark->group);
537 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
538 refcount_read(&mark->refcnt) < 1 +
539 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
540
541 spin_lock(&mark->lock);
542 /* something else already called this function on this mark */
543 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
544 spin_unlock(&mark->lock);
545 return;
546 }
547 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
548 list_del_init(&mark->g_list);
549 spin_unlock(&mark->lock);
550
551 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
552 fsnotify_put_mark(mark);
553}
554
555/*
556 * Free fsnotify mark. The mark is actually only marked as being freed. The
557 * freeing is actually happening only once last reference to the mark is
558 * dropped from a workqueue which first waits for srcu period end.
559 *
560 * Caller must have a reference to the mark or be protected by
561 * fsnotify_mark_srcu.
562 */
563void fsnotify_free_mark(struct fsnotify_mark *mark)
564{
565 struct fsnotify_group *group = mark->group;
566
567 spin_lock(&mark->lock);
568 /* something else already called this function on this mark */
569 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
570 spin_unlock(&mark->lock);
571 return;
572 }
573 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
574 spin_unlock(&mark->lock);
575
576 /*
577 * Some groups like to know that marks are being freed. This is a
578 * callback to the group function to let it know that this mark
579 * is being freed.
580 */
581 if (group->ops->freeing_mark)
582 group->ops->freeing_mark(mark, group);
583}
584
585void fsnotify_destroy_mark(struct fsnotify_mark *mark,
586 struct fsnotify_group *group)
587{
588 fsnotify_group_lock(group);
589 fsnotify_detach_mark(mark);
590 fsnotify_group_unlock(group);
591 fsnotify_free_mark(mark);
592}
593EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
594
595/*
596 * Sorting function for lists of fsnotify marks.
597 *
598 * Fanotify supports different notification classes (reflected as priority of
599 * notification group). Events shall be passed to notification groups in
600 * decreasing priority order. To achieve this marks in notification lists for
601 * inodes and vfsmounts are sorted so that priorities of corresponding groups
602 * are descending.
603 *
604 * Furthermore correct handling of the ignore mask requires processing inode
605 * and vfsmount marks of each group together. Using the group address as
606 * further sort criterion provides a unique sorting order and thus we can
607 * merge inode and vfsmount lists of marks in linear time and find groups
608 * present in both lists.
609 *
610 * A return value of 1 signifies that b has priority over a.
611 * A return value of 0 signifies that the two marks have to be handled together.
612 * A return value of -1 signifies that a has priority over b.
613 */
614int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
615{
616 if (a == b)
617 return 0;
618 if (!a)
619 return 1;
620 if (!b)
621 return -1;
622 if (a->priority < b->priority)
623 return 1;
624 if (a->priority > b->priority)
625 return -1;
626 if (a < b)
627 return 1;
628 return -1;
629}
630
631static int fsnotify_attach_info_to_sb(struct super_block *sb)
632{
633 struct fsnotify_sb_info *sbinfo;
634
635 /* sb info is freed on fsnotify_sb_delete() */
636 sbinfo = kzalloc(sizeof(*sbinfo), GFP_KERNEL);
637 if (!sbinfo)
638 return -ENOMEM;
639
640 /*
641 * cmpxchg() provides the barrier so that callers of fsnotify_sb_info()
642 * will observe an initialized structure
643 */
644 if (cmpxchg(&sb->s_fsnotify_info, NULL, sbinfo)) {
645 /* Someone else created sbinfo for us */
646 kfree(sbinfo);
647 }
648 return 0;
649}
650
651static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
652 void *obj, unsigned int obj_type)
653{
654 struct fsnotify_mark_connector *conn;
655
656 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
657 if (!conn)
658 return -ENOMEM;
659 spin_lock_init(&conn->lock);
660 INIT_HLIST_HEAD(&conn->list);
661 conn->flags = 0;
662 conn->prio = 0;
663 conn->type = obj_type;
664 conn->obj = obj;
665
666 /*
667 * cmpxchg() provides the barrier so that readers of *connp can see
668 * only initialized structure
669 */
670 if (cmpxchg(connp, NULL, conn)) {
671 /* Someone else created list structure for us */
672 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
673 }
674 return 0;
675}
676
677/*
678 * Get mark connector, make sure it is alive and return with its lock held.
679 * This is for users that get connector pointer from inode or mount. Users that
680 * hold reference to a mark on the list may directly lock connector->lock as
681 * they are sure list cannot go away under them.
682 */
683static struct fsnotify_mark_connector *fsnotify_grab_connector(
684 fsnotify_connp_t *connp)
685{
686 struct fsnotify_mark_connector *conn;
687 int idx;
688
689 idx = srcu_read_lock(&fsnotify_mark_srcu);
690 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
691 if (!conn)
692 goto out;
693 spin_lock(&conn->lock);
694 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
695 spin_unlock(&conn->lock);
696 srcu_read_unlock(&fsnotify_mark_srcu, idx);
697 return NULL;
698 }
699out:
700 srcu_read_unlock(&fsnotify_mark_srcu, idx);
701 return conn;
702}
703
704/*
705 * Add mark into proper place in given list of marks. These marks may be used
706 * for the fsnotify backend to determine which event types should be delivered
707 * to which group and for which inodes. These marks are ordered according to
708 * priority, highest number first, and then by the group's location in memory.
709 */
710static int fsnotify_add_mark_list(struct fsnotify_mark *mark, void *obj,
711 unsigned int obj_type, int add_flags)
712{
713 struct super_block *sb = fsnotify_object_sb(obj, obj_type);
714 struct fsnotify_mark *lmark, *last = NULL;
715 struct fsnotify_mark_connector *conn;
716 fsnotify_connp_t *connp;
717 int cmp;
718 int err = 0;
719
720 if (WARN_ON(!fsnotify_valid_obj_type(obj_type)))
721 return -EINVAL;
722
723 /*
724 * Attach the sb info before attaching a connector to any object on sb.
725 * The sb info will remain attached as long as sb lives.
726 */
727 if (!fsnotify_sb_info(sb)) {
728 err = fsnotify_attach_info_to_sb(sb);
729 if (err)
730 return err;
731 }
732
733 connp = fsnotify_object_connp(obj, obj_type);
734restart:
735 spin_lock(&mark->lock);
736 conn = fsnotify_grab_connector(connp);
737 if (!conn) {
738 spin_unlock(&mark->lock);
739 err = fsnotify_attach_connector_to_object(connp, obj, obj_type);
740 if (err)
741 return err;
742 goto restart;
743 }
744
745 /* is mark the first mark? */
746 if (hlist_empty(&conn->list)) {
747 hlist_add_head_rcu(&mark->obj_list, &conn->list);
748 goto added;
749 }
750
751 /* should mark be in the middle of the current list? */
752 hlist_for_each_entry(lmark, &conn->list, obj_list) {
753 last = lmark;
754
755 if ((lmark->group == mark->group) &&
756 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
757 !(mark->group->flags & FSNOTIFY_GROUP_DUPS)) {
758 err = -EEXIST;
759 goto out_err;
760 }
761
762 cmp = fsnotify_compare_groups(lmark->group, mark->group);
763 if (cmp >= 0) {
764 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
765 goto added;
766 }
767 }
768
769 BUG_ON(last == NULL);
770 /* mark should be the last entry. last is the current last entry */
771 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
772added:
773 fsnotify_update_sb_watchers(sb, conn);
774 /*
775 * Since connector is attached to object using cmpxchg() we are
776 * guaranteed that connector initialization is fully visible by anyone
777 * seeing mark->connector set.
778 */
779 WRITE_ONCE(mark->connector, conn);
780out_err:
781 spin_unlock(&conn->lock);
782 spin_unlock(&mark->lock);
783 return err;
784}
785
786/*
787 * Attach an initialized mark to a given group and fs object.
788 * These marks may be used for the fsnotify backend to determine which
789 * event types should be delivered to which group.
790 */
791int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
792 void *obj, unsigned int obj_type,
793 int add_flags)
794{
795 struct fsnotify_group *group = mark->group;
796 int ret = 0;
797
798 fsnotify_group_assert_locked(group);
799
800 /*
801 * LOCKING ORDER!!!!
802 * group->mark_mutex
803 * mark->lock
804 * mark->connector->lock
805 */
806 spin_lock(&mark->lock);
807 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
808
809 list_add(&mark->g_list, &group->marks_list);
810 fsnotify_get_mark(mark); /* for g_list */
811 spin_unlock(&mark->lock);
812
813 ret = fsnotify_add_mark_list(mark, obj, obj_type, add_flags);
814 if (ret)
815 goto err;
816
817 fsnotify_recalc_mask(mark->connector);
818
819 return ret;
820err:
821 spin_lock(&mark->lock);
822 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
823 FSNOTIFY_MARK_FLAG_ATTACHED);
824 list_del_init(&mark->g_list);
825 spin_unlock(&mark->lock);
826
827 fsnotify_put_mark(mark);
828 return ret;
829}
830
831int fsnotify_add_mark(struct fsnotify_mark *mark, void *obj,
832 unsigned int obj_type, int add_flags)
833{
834 int ret;
835 struct fsnotify_group *group = mark->group;
836
837 fsnotify_group_lock(group);
838 ret = fsnotify_add_mark_locked(mark, obj, obj_type, add_flags);
839 fsnotify_group_unlock(group);
840 return ret;
841}
842EXPORT_SYMBOL_GPL(fsnotify_add_mark);
843
844/*
845 * Given a list of marks, find the mark associated with given group. If found
846 * take a reference to that mark and return it, else return NULL.
847 */
848struct fsnotify_mark *fsnotify_find_mark(void *obj, unsigned int obj_type,
849 struct fsnotify_group *group)
850{
851 fsnotify_connp_t *connp = fsnotify_object_connp(obj, obj_type);
852 struct fsnotify_mark_connector *conn;
853 struct fsnotify_mark *mark;
854
855 if (!connp)
856 return NULL;
857
858 conn = fsnotify_grab_connector(connp);
859 if (!conn)
860 return NULL;
861
862 hlist_for_each_entry(mark, &conn->list, obj_list) {
863 if (mark->group == group &&
864 (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
865 fsnotify_get_mark(mark);
866 spin_unlock(&conn->lock);
867 return mark;
868 }
869 }
870 spin_unlock(&conn->lock);
871 return NULL;
872}
873EXPORT_SYMBOL_GPL(fsnotify_find_mark);
874
875/* Clear any marks in a group with given type mask */
876void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
877 unsigned int obj_type)
878{
879 struct fsnotify_mark *lmark, *mark;
880 LIST_HEAD(to_free);
881 struct list_head *head = &to_free;
882
883 /* Skip selection step if we want to clear all marks. */
884 if (obj_type == FSNOTIFY_OBJ_TYPE_ANY) {
885 head = &group->marks_list;
886 goto clear;
887 }
888 /*
889 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
890 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
891 * to_free list so we have to use mark_mutex even when accessing that
892 * list. And freeing mark requires us to drop mark_mutex. So we can
893 * reliably free only the first mark in the list. That's why we first
894 * move marks to free to to_free list in one go and then free marks in
895 * to_free list one by one.
896 */
897 fsnotify_group_lock(group);
898 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
899 if (mark->connector->type == obj_type)
900 list_move(&mark->g_list, &to_free);
901 }
902 fsnotify_group_unlock(group);
903
904clear:
905 while (1) {
906 fsnotify_group_lock(group);
907 if (list_empty(head)) {
908 fsnotify_group_unlock(group);
909 break;
910 }
911 mark = list_first_entry(head, struct fsnotify_mark, g_list);
912 fsnotify_get_mark(mark);
913 fsnotify_detach_mark(mark);
914 fsnotify_group_unlock(group);
915 fsnotify_free_mark(mark);
916 fsnotify_put_mark(mark);
917 }
918}
919
920/* Destroy all marks attached to an object via connector */
921void fsnotify_destroy_marks(fsnotify_connp_t *connp)
922{
923 struct fsnotify_mark_connector *conn;
924 struct fsnotify_mark *mark, *old_mark = NULL;
925 void *objp;
926 unsigned int type;
927
928 conn = fsnotify_grab_connector(connp);
929 if (!conn)
930 return;
931 /*
932 * We have to be careful since we can race with e.g.
933 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
934 * list can get modified. However we are holding mark reference and
935 * thus our mark cannot be removed from obj_list so we can continue
936 * iteration after regaining conn->lock.
937 */
938 hlist_for_each_entry(mark, &conn->list, obj_list) {
939 fsnotify_get_mark(mark);
940 spin_unlock(&conn->lock);
941 if (old_mark)
942 fsnotify_put_mark(old_mark);
943 old_mark = mark;
944 fsnotify_destroy_mark(mark, mark->group);
945 spin_lock(&conn->lock);
946 }
947 /*
948 * Detach list from object now so that we don't pin inode until all
949 * mark references get dropped. It would lead to strange results such
950 * as delaying inode deletion or blocking unmount.
951 */
952 objp = fsnotify_detach_connector_from_object(conn, &type);
953 spin_unlock(&conn->lock);
954 if (old_mark)
955 fsnotify_put_mark(old_mark);
956 fsnotify_drop_object(type, objp);
957}
958
959/*
960 * Nothing fancy, just initialize lists and locks and counters.
961 */
962void fsnotify_init_mark(struct fsnotify_mark *mark,
963 struct fsnotify_group *group)
964{
965 memset(mark, 0, sizeof(*mark));
966 spin_lock_init(&mark->lock);
967 refcount_set(&mark->refcnt, 1);
968 fsnotify_get_group(group);
969 mark->group = group;
970 WRITE_ONCE(mark->connector, NULL);
971}
972EXPORT_SYMBOL_GPL(fsnotify_init_mark);
973
974/*
975 * Destroy all marks in destroy_list, waits for SRCU period to finish before
976 * actually freeing marks.
977 */
978static void fsnotify_mark_destroy_workfn(struct work_struct *work)
979{
980 struct fsnotify_mark *mark, *next;
981 struct list_head private_destroy_list;
982
983 spin_lock(&destroy_lock);
984 /* exchange the list head */
985 list_replace_init(&destroy_list, &private_destroy_list);
986 spin_unlock(&destroy_lock);
987
988 synchronize_srcu(&fsnotify_mark_srcu);
989
990 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
991 list_del_init(&mark->g_list);
992 fsnotify_final_mark_destroy(mark);
993 }
994}
995
996/* Wait for all marks queued for destruction to be actually destroyed */
997void fsnotify_wait_marks_destroyed(void)
998{
999 flush_delayed_work(&reaper_work);
1000}
1001EXPORT_SYMBOL_GPL(fsnotify_wait_marks_destroyed);