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 __u32 *fsnotify_conn_mask_p(struct fsnotify_mark_connector *conn)
101{
102 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
103 return &fsnotify_conn_inode(conn)->i_fsnotify_mask;
104 else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT)
105 return &fsnotify_conn_mount(conn)->mnt_fsnotify_mask;
106 else if (conn->type == FSNOTIFY_OBJ_TYPE_SB)
107 return &fsnotify_conn_sb(conn)->s_fsnotify_mask;
108 return NULL;
109}
110
111__u32 fsnotify_conn_mask(struct fsnotify_mark_connector *conn)
112{
113 if (WARN_ON(!fsnotify_valid_obj_type(conn->type)))
114 return 0;
115
116 return *fsnotify_conn_mask_p(conn);
117}
118
119static void __fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
120{
121 u32 new_mask = 0;
122 struct fsnotify_mark *mark;
123
124 assert_spin_locked(&conn->lock);
125 /* We can get detached connector here when inode is getting unlinked. */
126 if (!fsnotify_valid_obj_type(conn->type))
127 return;
128 hlist_for_each_entry(mark, &conn->list, obj_list) {
129 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)
130 new_mask |= mark->mask;
131 }
132 *fsnotify_conn_mask_p(conn) = new_mask;
133}
134
135/*
136 * Calculate mask of events for a list of marks. The caller must make sure
137 * connector and connector->obj cannot disappear under us. Callers achieve
138 * this by holding a mark->lock or mark->group->mark_mutex for a mark on this
139 * list.
140 */
141void fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
142{
143 if (!conn)
144 return;
145
146 spin_lock(&conn->lock);
147 __fsnotify_recalc_mask(conn);
148 spin_unlock(&conn->lock);
149 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
150 __fsnotify_update_child_dentry_flags(
151 fsnotify_conn_inode(conn));
152}
153
154/* Free all connectors queued for freeing once SRCU period ends */
155static void fsnotify_connector_destroy_workfn(struct work_struct *work)
156{
157 struct fsnotify_mark_connector *conn, *free;
158
159 spin_lock(&destroy_lock);
160 conn = connector_destroy_list;
161 connector_destroy_list = NULL;
162 spin_unlock(&destroy_lock);
163
164 synchronize_srcu(&fsnotify_mark_srcu);
165 while (conn) {
166 free = conn;
167 conn = conn->destroy_next;
168 kmem_cache_free(fsnotify_mark_connector_cachep, free);
169 }
170}
171
172static void *fsnotify_detach_connector_from_object(
173 struct fsnotify_mark_connector *conn,
174 unsigned int *type)
175{
176 struct inode *inode = NULL;
177
178 *type = conn->type;
179 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED)
180 return NULL;
181
182 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE) {
183 inode = fsnotify_conn_inode(conn);
184 inode->i_fsnotify_mask = 0;
185 atomic_long_inc(&inode->i_sb->s_fsnotify_inode_refs);
186 } else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT) {
187 fsnotify_conn_mount(conn)->mnt_fsnotify_mask = 0;
188 } else if (conn->type == FSNOTIFY_OBJ_TYPE_SB) {
189 fsnotify_conn_sb(conn)->s_fsnotify_mask = 0;
190 }
191
192 rcu_assign_pointer(*(conn->obj), NULL);
193 conn->obj = NULL;
194 conn->type = FSNOTIFY_OBJ_TYPE_DETACHED;
195
196 return inode;
197}
198
199static void fsnotify_final_mark_destroy(struct fsnotify_mark *mark)
200{
201 struct fsnotify_group *group = mark->group;
202
203 if (WARN_ON_ONCE(!group))
204 return;
205 group->ops->free_mark(mark);
206 fsnotify_put_group(group);
207}
208
209/* Drop object reference originally held by a connector */
210static void fsnotify_drop_object(unsigned int type, void *objp)
211{
212 struct inode *inode;
213 struct super_block *sb;
214
215 if (!objp)
216 return;
217 /* Currently only inode references are passed to be dropped */
218 if (WARN_ON_ONCE(type != FSNOTIFY_OBJ_TYPE_INODE))
219 return;
220 inode = objp;
221 sb = inode->i_sb;
222 iput(inode);
223 if (atomic_long_dec_and_test(&sb->s_fsnotify_inode_refs))
224 wake_up_var(&sb->s_fsnotify_inode_refs);
225}
226
227void fsnotify_put_mark(struct fsnotify_mark *mark)
228{
229 struct fsnotify_mark_connector *conn = READ_ONCE(mark->connector);
230 void *objp = NULL;
231 unsigned int type = FSNOTIFY_OBJ_TYPE_DETACHED;
232 bool free_conn = false;
233
234 /* Catch marks that were actually never attached to object */
235 if (!conn) {
236 if (refcount_dec_and_test(&mark->refcnt))
237 fsnotify_final_mark_destroy(mark);
238 return;
239 }
240
241 /*
242 * We have to be careful so that traversals of obj_list under lock can
243 * safely grab mark reference.
244 */
245 if (!refcount_dec_and_lock(&mark->refcnt, &conn->lock))
246 return;
247
248 hlist_del_init_rcu(&mark->obj_list);
249 if (hlist_empty(&conn->list)) {
250 objp = fsnotify_detach_connector_from_object(conn, &type);
251 free_conn = true;
252 } else {
253 __fsnotify_recalc_mask(conn);
254 }
255 WRITE_ONCE(mark->connector, NULL);
256 spin_unlock(&conn->lock);
257
258 fsnotify_drop_object(type, objp);
259
260 if (free_conn) {
261 spin_lock(&destroy_lock);
262 conn->destroy_next = connector_destroy_list;
263 connector_destroy_list = conn;
264 spin_unlock(&destroy_lock);
265 queue_work(system_unbound_wq, &connector_reaper_work);
266 }
267 /*
268 * Note that we didn't update flags telling whether inode cares about
269 * what's happening with children. We update these flags from
270 * __fsnotify_parent() lazily when next event happens on one of our
271 * children.
272 */
273 spin_lock(&destroy_lock);
274 list_add(&mark->g_list, &destroy_list);
275 spin_unlock(&destroy_lock);
276 queue_delayed_work(system_unbound_wq, &reaper_work,
277 FSNOTIFY_REAPER_DELAY);
278}
279EXPORT_SYMBOL_GPL(fsnotify_put_mark);
280
281/*
282 * Get mark reference when we found the mark via lockless traversal of object
283 * list. Mark can be already removed from the list by now and on its way to be
284 * destroyed once SRCU period ends.
285 *
286 * Also pin the group so it doesn't disappear under us.
287 */
288static bool fsnotify_get_mark_safe(struct fsnotify_mark *mark)
289{
290 if (!mark)
291 return true;
292
293 if (refcount_inc_not_zero(&mark->refcnt)) {
294 spin_lock(&mark->lock);
295 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) {
296 /* mark is attached, group is still alive then */
297 atomic_inc(&mark->group->user_waits);
298 spin_unlock(&mark->lock);
299 return true;
300 }
301 spin_unlock(&mark->lock);
302 fsnotify_put_mark(mark);
303 }
304 return false;
305}
306
307/*
308 * Puts marks and wakes up group destruction if necessary.
309 *
310 * Pairs with fsnotify_get_mark_safe()
311 */
312static void fsnotify_put_mark_wake(struct fsnotify_mark *mark)
313{
314 if (mark) {
315 struct fsnotify_group *group = mark->group;
316
317 fsnotify_put_mark(mark);
318 /*
319 * We abuse notification_waitq on group shutdown for waiting for
320 * all marks pinned when waiting for userspace.
321 */
322 if (atomic_dec_and_test(&group->user_waits) && group->shutdown)
323 wake_up(&group->notification_waitq);
324 }
325}
326
327bool fsnotify_prepare_user_wait(struct fsnotify_iter_info *iter_info)
328{
329 int type;
330
331 fsnotify_foreach_obj_type(type) {
332 /* This can fail if mark is being removed */
333 if (!fsnotify_get_mark_safe(iter_info->marks[type]))
334 goto fail;
335 }
336
337 /*
338 * Now that both marks are pinned by refcount in the inode / vfsmount
339 * lists, we can drop SRCU lock, and safely resume the list iteration
340 * once userspace returns.
341 */
342 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
343
344 return true;
345
346fail:
347 for (type--; type >= 0; type--)
348 fsnotify_put_mark_wake(iter_info->marks[type]);
349 return false;
350}
351
352void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
353{
354 int type;
355
356 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
357 fsnotify_foreach_obj_type(type)
358 fsnotify_put_mark_wake(iter_info->marks[type]);
359}
360
361/*
362 * Mark mark as detached, remove it from group list. Mark still stays in object
363 * list until its last reference is dropped. Note that we rely on mark being
364 * removed from group list before corresponding reference to it is dropped. In
365 * particular we rely on mark->connector being valid while we hold
366 * group->mark_mutex if we found the mark through g_list.
367 *
368 * Must be called with group->mark_mutex held. The caller must either hold
369 * reference to the mark or be protected by fsnotify_mark_srcu.
370 */
371void fsnotify_detach_mark(struct fsnotify_mark *mark)
372{
373 struct fsnotify_group *group = mark->group;
374
375 WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
376 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
377 refcount_read(&mark->refcnt) < 1 +
378 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
379
380 spin_lock(&mark->lock);
381 /* something else already called this function on this mark */
382 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
383 spin_unlock(&mark->lock);
384 return;
385 }
386 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
387 list_del_init(&mark->g_list);
388 spin_unlock(&mark->lock);
389
390 atomic_dec(&group->num_marks);
391
392 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
393 fsnotify_put_mark(mark);
394}
395
396/*
397 * Free fsnotify mark. The mark is actually only marked as being freed. The
398 * freeing is actually happening only once last reference to the mark is
399 * dropped from a workqueue which first waits for srcu period end.
400 *
401 * Caller must have a reference to the mark or be protected by
402 * fsnotify_mark_srcu.
403 */
404void fsnotify_free_mark(struct fsnotify_mark *mark)
405{
406 struct fsnotify_group *group = mark->group;
407
408 spin_lock(&mark->lock);
409 /* something else already called this function on this mark */
410 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
411 spin_unlock(&mark->lock);
412 return;
413 }
414 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
415 spin_unlock(&mark->lock);
416
417 /*
418 * Some groups like to know that marks are being freed. This is a
419 * callback to the group function to let it know that this mark
420 * is being freed.
421 */
422 if (group->ops->freeing_mark)
423 group->ops->freeing_mark(mark, group);
424}
425
426void fsnotify_destroy_mark(struct fsnotify_mark *mark,
427 struct fsnotify_group *group)
428{
429 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
430 fsnotify_detach_mark(mark);
431 mutex_unlock(&group->mark_mutex);
432 fsnotify_free_mark(mark);
433}
434EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
435
436/*
437 * Sorting function for lists of fsnotify marks.
438 *
439 * Fanotify supports different notification classes (reflected as priority of
440 * notification group). Events shall be passed to notification groups in
441 * decreasing priority order. To achieve this marks in notification lists for
442 * inodes and vfsmounts are sorted so that priorities of corresponding groups
443 * are descending.
444 *
445 * Furthermore correct handling of the ignore mask requires processing inode
446 * and vfsmount marks of each group together. Using the group address as
447 * further sort criterion provides a unique sorting order and thus we can
448 * merge inode and vfsmount lists of marks in linear time and find groups
449 * present in both lists.
450 *
451 * A return value of 1 signifies that b has priority over a.
452 * A return value of 0 signifies that the two marks have to be handled together.
453 * A return value of -1 signifies that a has priority over b.
454 */
455int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
456{
457 if (a == b)
458 return 0;
459 if (!a)
460 return 1;
461 if (!b)
462 return -1;
463 if (a->priority < b->priority)
464 return 1;
465 if (a->priority > b->priority)
466 return -1;
467 if (a < b)
468 return 1;
469 return -1;
470}
471
472static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
473 unsigned int type,
474 __kernel_fsid_t *fsid)
475{
476 struct inode *inode = NULL;
477 struct fsnotify_mark_connector *conn;
478
479 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
480 if (!conn)
481 return -ENOMEM;
482 spin_lock_init(&conn->lock);
483 INIT_HLIST_HEAD(&conn->list);
484 conn->type = type;
485 conn->obj = connp;
486 /* Cache fsid of filesystem containing the object */
487 if (fsid) {
488 conn->fsid = *fsid;
489 conn->flags = FSNOTIFY_CONN_FLAG_HAS_FSID;
490 } else {
491 conn->fsid.val[0] = conn->fsid.val[1] = 0;
492 conn->flags = 0;
493 }
494 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
495 inode = igrab(fsnotify_conn_inode(conn));
496 /*
497 * cmpxchg() provides the barrier so that readers of *connp can see
498 * only initialized structure
499 */
500 if (cmpxchg(connp, NULL, conn)) {
501 /* Someone else created list structure for us */
502 if (inode)
503 iput(inode);
504 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
505 }
506
507 return 0;
508}
509
510/*
511 * Get mark connector, make sure it is alive and return with its lock held.
512 * This is for users that get connector pointer from inode or mount. Users that
513 * hold reference to a mark on the list may directly lock connector->lock as
514 * they are sure list cannot go away under them.
515 */
516static struct fsnotify_mark_connector *fsnotify_grab_connector(
517 fsnotify_connp_t *connp)
518{
519 struct fsnotify_mark_connector *conn;
520 int idx;
521
522 idx = srcu_read_lock(&fsnotify_mark_srcu);
523 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
524 if (!conn)
525 goto out;
526 spin_lock(&conn->lock);
527 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
528 spin_unlock(&conn->lock);
529 srcu_read_unlock(&fsnotify_mark_srcu, idx);
530 return NULL;
531 }
532out:
533 srcu_read_unlock(&fsnotify_mark_srcu, idx);
534 return conn;
535}
536
537/*
538 * Add mark into proper place in given list of marks. These marks may be used
539 * for the fsnotify backend to determine which event types should be delivered
540 * to which group and for which inodes. These marks are ordered according to
541 * priority, highest number first, and then by the group's location in memory.
542 */
543static int fsnotify_add_mark_list(struct fsnotify_mark *mark,
544 fsnotify_connp_t *connp, unsigned int type,
545 int allow_dups, __kernel_fsid_t *fsid)
546{
547 struct fsnotify_mark *lmark, *last = NULL;
548 struct fsnotify_mark_connector *conn;
549 int cmp;
550 int err = 0;
551
552 if (WARN_ON(!fsnotify_valid_obj_type(type)))
553 return -EINVAL;
554
555 /* Backend is expected to check for zero fsid (e.g. tmpfs) */
556 if (fsid && WARN_ON_ONCE(!fsid->val[0] && !fsid->val[1]))
557 return -ENODEV;
558
559restart:
560 spin_lock(&mark->lock);
561 conn = fsnotify_grab_connector(connp);
562 if (!conn) {
563 spin_unlock(&mark->lock);
564 err = fsnotify_attach_connector_to_object(connp, type, fsid);
565 if (err)
566 return err;
567 goto restart;
568 } else if (fsid && !(conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID)) {
569 conn->fsid = *fsid;
570 /* Pairs with smp_rmb() in fanotify_get_fsid() */
571 smp_wmb();
572 conn->flags |= FSNOTIFY_CONN_FLAG_HAS_FSID;
573 } else if (fsid && (conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID) &&
574 (fsid->val[0] != conn->fsid.val[0] ||
575 fsid->val[1] != conn->fsid.val[1])) {
576 /*
577 * Backend is expected to check for non uniform fsid
578 * (e.g. btrfs), but maybe we missed something?
579 * Only allow setting conn->fsid once to non zero fsid.
580 * inotify and non-fid fanotify groups do not set nor test
581 * conn->fsid.
582 */
583 pr_warn_ratelimited("%s: fsid mismatch on object of type %u: "
584 "%x.%x != %x.%x\n", __func__, conn->type,
585 fsid->val[0], fsid->val[1],
586 conn->fsid.val[0], conn->fsid.val[1]);
587 err = -EXDEV;
588 goto out_err;
589 }
590
591 /* is mark the first mark? */
592 if (hlist_empty(&conn->list)) {
593 hlist_add_head_rcu(&mark->obj_list, &conn->list);
594 goto added;
595 }
596
597 /* should mark be in the middle of the current list? */
598 hlist_for_each_entry(lmark, &conn->list, obj_list) {
599 last = lmark;
600
601 if ((lmark->group == mark->group) &&
602 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
603 !allow_dups) {
604 err = -EEXIST;
605 goto out_err;
606 }
607
608 cmp = fsnotify_compare_groups(lmark->group, mark->group);
609 if (cmp >= 0) {
610 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
611 goto added;
612 }
613 }
614
615 BUG_ON(last == NULL);
616 /* mark should be the last entry. last is the current last entry */
617 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
618added:
619 /*
620 * Since connector is attached to object using cmpxchg() we are
621 * guaranteed that connector initialization is fully visible by anyone
622 * seeing mark->connector set.
623 */
624 WRITE_ONCE(mark->connector, conn);
625out_err:
626 spin_unlock(&conn->lock);
627 spin_unlock(&mark->lock);
628 return err;
629}
630
631/*
632 * Attach an initialized mark to a given group and fs object.
633 * These marks may be used for the fsnotify backend to determine which
634 * event types should be delivered to which group.
635 */
636int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
637 fsnotify_connp_t *connp, unsigned int type,
638 int allow_dups, __kernel_fsid_t *fsid)
639{
640 struct fsnotify_group *group = mark->group;
641 int ret = 0;
642
643 BUG_ON(!mutex_is_locked(&group->mark_mutex));
644
645 /*
646 * LOCKING ORDER!!!!
647 * group->mark_mutex
648 * mark->lock
649 * mark->connector->lock
650 */
651 spin_lock(&mark->lock);
652 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
653
654 list_add(&mark->g_list, &group->marks_list);
655 atomic_inc(&group->num_marks);
656 fsnotify_get_mark(mark); /* for g_list */
657 spin_unlock(&mark->lock);
658
659 ret = fsnotify_add_mark_list(mark, connp, type, allow_dups, fsid);
660 if (ret)
661 goto err;
662
663 if (mark->mask)
664 fsnotify_recalc_mask(mark->connector);
665
666 return ret;
667err:
668 spin_lock(&mark->lock);
669 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
670 FSNOTIFY_MARK_FLAG_ATTACHED);
671 list_del_init(&mark->g_list);
672 spin_unlock(&mark->lock);
673 atomic_dec(&group->num_marks);
674
675 fsnotify_put_mark(mark);
676 return ret;
677}
678
679int fsnotify_add_mark(struct fsnotify_mark *mark, fsnotify_connp_t *connp,
680 unsigned int type, int allow_dups, __kernel_fsid_t *fsid)
681{
682 int ret;
683 struct fsnotify_group *group = mark->group;
684
685 mutex_lock(&group->mark_mutex);
686 ret = fsnotify_add_mark_locked(mark, connp, type, allow_dups, fsid);
687 mutex_unlock(&group->mark_mutex);
688 return ret;
689}
690EXPORT_SYMBOL_GPL(fsnotify_add_mark);
691
692/*
693 * Given a list of marks, find the mark associated with given group. If found
694 * take a reference to that mark and return it, else return NULL.
695 */
696struct fsnotify_mark *fsnotify_find_mark(fsnotify_connp_t *connp,
697 struct fsnotify_group *group)
698{
699 struct fsnotify_mark_connector *conn;
700 struct fsnotify_mark *mark;
701
702 conn = fsnotify_grab_connector(connp);
703 if (!conn)
704 return NULL;
705
706 hlist_for_each_entry(mark, &conn->list, obj_list) {
707 if (mark->group == group &&
708 (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
709 fsnotify_get_mark(mark);
710 spin_unlock(&conn->lock);
711 return mark;
712 }
713 }
714 spin_unlock(&conn->lock);
715 return NULL;
716}
717EXPORT_SYMBOL_GPL(fsnotify_find_mark);
718
719/* Clear any marks in a group with given type mask */
720void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
721 unsigned int type_mask)
722{
723 struct fsnotify_mark *lmark, *mark;
724 LIST_HEAD(to_free);
725 struct list_head *head = &to_free;
726
727 /* Skip selection step if we want to clear all marks. */
728 if (type_mask == FSNOTIFY_OBJ_ALL_TYPES_MASK) {
729 head = &group->marks_list;
730 goto clear;
731 }
732 /*
733 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
734 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
735 * to_free list so we have to use mark_mutex even when accessing that
736 * list. And freeing mark requires us to drop mark_mutex. So we can
737 * reliably free only the first mark in the list. That's why we first
738 * move marks to free to to_free list in one go and then free marks in
739 * to_free list one by one.
740 */
741 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
742 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
743 if ((1U << mark->connector->type) & type_mask)
744 list_move(&mark->g_list, &to_free);
745 }
746 mutex_unlock(&group->mark_mutex);
747
748clear:
749 while (1) {
750 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
751 if (list_empty(head)) {
752 mutex_unlock(&group->mark_mutex);
753 break;
754 }
755 mark = list_first_entry(head, struct fsnotify_mark, g_list);
756 fsnotify_get_mark(mark);
757 fsnotify_detach_mark(mark);
758 mutex_unlock(&group->mark_mutex);
759 fsnotify_free_mark(mark);
760 fsnotify_put_mark(mark);
761 }
762}
763
764/* Destroy all marks attached to an object via connector */
765void fsnotify_destroy_marks(fsnotify_connp_t *connp)
766{
767 struct fsnotify_mark_connector *conn;
768 struct fsnotify_mark *mark, *old_mark = NULL;
769 void *objp;
770 unsigned int type;
771
772 conn = fsnotify_grab_connector(connp);
773 if (!conn)
774 return;
775 /*
776 * We have to be careful since we can race with e.g.
777 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
778 * list can get modified. However we are holding mark reference and
779 * thus our mark cannot be removed from obj_list so we can continue
780 * iteration after regaining conn->lock.
781 */
782 hlist_for_each_entry(mark, &conn->list, obj_list) {
783 fsnotify_get_mark(mark);
784 spin_unlock(&conn->lock);
785 if (old_mark)
786 fsnotify_put_mark(old_mark);
787 old_mark = mark;
788 fsnotify_destroy_mark(mark, mark->group);
789 spin_lock(&conn->lock);
790 }
791 /*
792 * Detach list from object now so that we don't pin inode until all
793 * mark references get dropped. It would lead to strange results such
794 * as delaying inode deletion or blocking unmount.
795 */
796 objp = fsnotify_detach_connector_from_object(conn, &type);
797 spin_unlock(&conn->lock);
798 if (old_mark)
799 fsnotify_put_mark(old_mark);
800 fsnotify_drop_object(type, objp);
801}
802
803/*
804 * Nothing fancy, just initialize lists and locks and counters.
805 */
806void fsnotify_init_mark(struct fsnotify_mark *mark,
807 struct fsnotify_group *group)
808{
809 memset(mark, 0, sizeof(*mark));
810 spin_lock_init(&mark->lock);
811 refcount_set(&mark->refcnt, 1);
812 fsnotify_get_group(group);
813 mark->group = group;
814 WRITE_ONCE(mark->connector, NULL);
815}
816EXPORT_SYMBOL_GPL(fsnotify_init_mark);
817
818/*
819 * Destroy all marks in destroy_list, waits for SRCU period to finish before
820 * actually freeing marks.
821 */
822static void fsnotify_mark_destroy_workfn(struct work_struct *work)
823{
824 struct fsnotify_mark *mark, *next;
825 struct list_head private_destroy_list;
826
827 spin_lock(&destroy_lock);
828 /* exchange the list head */
829 list_replace_init(&destroy_list, &private_destroy_list);
830 spin_unlock(&destroy_lock);
831
832 synchronize_srcu(&fsnotify_mark_srcu);
833
834 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
835 list_del_init(&mark->g_list);
836 fsnotify_final_mark_destroy(mark);
837 }
838}
839
840/* Wait for all marks queued for destruction to be actually destroyed */
841void fsnotify_wait_marks_destroyed(void)
842{
843 flush_delayed_work(&reaper_work);
844}
845EXPORT_SYMBOL_GPL(fsnotify_wait_marks_destroyed);