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_get_inode_ref(struct inode *inode)
120{
121 ihold(inode);
122 atomic_long_inc(&inode->i_sb->s_fsnotify_connectors);
123}
124
125/*
126 * Grab or drop inode reference for the connector if needed.
127 *
128 * When it's time to drop the reference, we only clear the HAS_IREF flag and
129 * return the inode object. fsnotify_drop_object() will be resonsible for doing
130 * iput() outside of spinlocks. This happens when last mark that wanted iref is
131 * detached.
132 */
133static struct inode *fsnotify_update_iref(struct fsnotify_mark_connector *conn,
134 bool want_iref)
135{
136 bool has_iref = conn->flags & FSNOTIFY_CONN_FLAG_HAS_IREF;
137 struct inode *inode = NULL;
138
139 if (conn->type != FSNOTIFY_OBJ_TYPE_INODE ||
140 want_iref == has_iref)
141 return NULL;
142
143 if (want_iref) {
144 /* Pin inode if any mark wants inode refcount held */
145 fsnotify_get_inode_ref(fsnotify_conn_inode(conn));
146 conn->flags |= FSNOTIFY_CONN_FLAG_HAS_IREF;
147 } else {
148 /* Unpin inode after detach of last mark that wanted iref */
149 inode = fsnotify_conn_inode(conn);
150 conn->flags &= ~FSNOTIFY_CONN_FLAG_HAS_IREF;
151 }
152
153 return inode;
154}
155
156static void *__fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
157{
158 u32 new_mask = 0;
159 bool want_iref = false;
160 struct fsnotify_mark *mark;
161
162 assert_spin_locked(&conn->lock);
163 /* We can get detached connector here when inode is getting unlinked. */
164 if (!fsnotify_valid_obj_type(conn->type))
165 return NULL;
166 hlist_for_each_entry(mark, &conn->list, obj_list) {
167 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED))
168 continue;
169 new_mask |= fsnotify_calc_mask(mark);
170 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE &&
171 !(mark->flags & FSNOTIFY_MARK_FLAG_NO_IREF))
172 want_iref = true;
173 }
174 *fsnotify_conn_mask_p(conn) = new_mask;
175
176 return fsnotify_update_iref(conn, want_iref);
177}
178
179/*
180 * Calculate mask of events for a list of marks. The caller must make sure
181 * connector and connector->obj cannot disappear under us. Callers achieve
182 * this by holding a mark->lock or mark->group->mark_mutex for a mark on this
183 * list.
184 */
185void fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
186{
187 if (!conn)
188 return;
189
190 spin_lock(&conn->lock);
191 __fsnotify_recalc_mask(conn);
192 spin_unlock(&conn->lock);
193 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
194 __fsnotify_update_child_dentry_flags(
195 fsnotify_conn_inode(conn));
196}
197
198/* Free all connectors queued for freeing once SRCU period ends */
199static void fsnotify_connector_destroy_workfn(struct work_struct *work)
200{
201 struct fsnotify_mark_connector *conn, *free;
202
203 spin_lock(&destroy_lock);
204 conn = connector_destroy_list;
205 connector_destroy_list = NULL;
206 spin_unlock(&destroy_lock);
207
208 synchronize_srcu(&fsnotify_mark_srcu);
209 while (conn) {
210 free = conn;
211 conn = conn->destroy_next;
212 kmem_cache_free(fsnotify_mark_connector_cachep, free);
213 }
214}
215
216static void fsnotify_put_inode_ref(struct inode *inode)
217{
218 struct super_block *sb = inode->i_sb;
219
220 iput(inode);
221 if (atomic_long_dec_and_test(&sb->s_fsnotify_connectors))
222 wake_up_var(&sb->s_fsnotify_connectors);
223}
224
225static void fsnotify_get_sb_connectors(struct fsnotify_mark_connector *conn)
226{
227 struct super_block *sb = fsnotify_connector_sb(conn);
228
229 if (sb)
230 atomic_long_inc(&sb->s_fsnotify_connectors);
231}
232
233static void fsnotify_put_sb_connectors(struct fsnotify_mark_connector *conn)
234{
235 struct super_block *sb = fsnotify_connector_sb(conn);
236
237 if (sb && atomic_long_dec_and_test(&sb->s_fsnotify_connectors))
238 wake_up_var(&sb->s_fsnotify_connectors);
239}
240
241static void *fsnotify_detach_connector_from_object(
242 struct fsnotify_mark_connector *conn,
243 unsigned int *type)
244{
245 struct inode *inode = NULL;
246
247 *type = conn->type;
248 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED)
249 return NULL;
250
251 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE) {
252 inode = fsnotify_conn_inode(conn);
253 inode->i_fsnotify_mask = 0;
254
255 /* Unpin inode when detaching from connector */
256 if (!(conn->flags & FSNOTIFY_CONN_FLAG_HAS_IREF))
257 inode = NULL;
258 } else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT) {
259 fsnotify_conn_mount(conn)->mnt_fsnotify_mask = 0;
260 } else if (conn->type == FSNOTIFY_OBJ_TYPE_SB) {
261 fsnotify_conn_sb(conn)->s_fsnotify_mask = 0;
262 }
263
264 fsnotify_put_sb_connectors(conn);
265 rcu_assign_pointer(*(conn->obj), NULL);
266 conn->obj = NULL;
267 conn->type = FSNOTIFY_OBJ_TYPE_DETACHED;
268
269 return inode;
270}
271
272static void fsnotify_final_mark_destroy(struct fsnotify_mark *mark)
273{
274 struct fsnotify_group *group = mark->group;
275
276 if (WARN_ON_ONCE(!group))
277 return;
278 group->ops->free_mark(mark);
279 fsnotify_put_group(group);
280}
281
282/* Drop object reference originally held by a connector */
283static void fsnotify_drop_object(unsigned int type, void *objp)
284{
285 if (!objp)
286 return;
287 /* Currently only inode references are passed to be dropped */
288 if (WARN_ON_ONCE(type != FSNOTIFY_OBJ_TYPE_INODE))
289 return;
290 fsnotify_put_inode_ref(objp);
291}
292
293void fsnotify_put_mark(struct fsnotify_mark *mark)
294{
295 struct fsnotify_mark_connector *conn = READ_ONCE(mark->connector);
296 void *objp = NULL;
297 unsigned int type = FSNOTIFY_OBJ_TYPE_DETACHED;
298 bool free_conn = false;
299
300 /* Catch marks that were actually never attached to object */
301 if (!conn) {
302 if (refcount_dec_and_test(&mark->refcnt))
303 fsnotify_final_mark_destroy(mark);
304 return;
305 }
306
307 /*
308 * We have to be careful so that traversals of obj_list under lock can
309 * safely grab mark reference.
310 */
311 if (!refcount_dec_and_lock(&mark->refcnt, &conn->lock))
312 return;
313
314 hlist_del_init_rcu(&mark->obj_list);
315 if (hlist_empty(&conn->list)) {
316 objp = fsnotify_detach_connector_from_object(conn, &type);
317 free_conn = true;
318 } else {
319 objp = __fsnotify_recalc_mask(conn);
320 type = conn->type;
321 }
322 WRITE_ONCE(mark->connector, NULL);
323 spin_unlock(&conn->lock);
324
325 fsnotify_drop_object(type, objp);
326
327 if (free_conn) {
328 spin_lock(&destroy_lock);
329 conn->destroy_next = connector_destroy_list;
330 connector_destroy_list = conn;
331 spin_unlock(&destroy_lock);
332 queue_work(system_unbound_wq, &connector_reaper_work);
333 }
334 /*
335 * Note that we didn't update flags telling whether inode cares about
336 * what's happening with children. We update these flags from
337 * __fsnotify_parent() lazily when next event happens on one of our
338 * children.
339 */
340 spin_lock(&destroy_lock);
341 list_add(&mark->g_list, &destroy_list);
342 spin_unlock(&destroy_lock);
343 queue_delayed_work(system_unbound_wq, &reaper_work,
344 FSNOTIFY_REAPER_DELAY);
345}
346EXPORT_SYMBOL_GPL(fsnotify_put_mark);
347
348/*
349 * Get mark reference when we found the mark via lockless traversal of object
350 * list. Mark can be already removed from the list by now and on its way to be
351 * destroyed once SRCU period ends.
352 *
353 * Also pin the group so it doesn't disappear under us.
354 */
355static bool fsnotify_get_mark_safe(struct fsnotify_mark *mark)
356{
357 if (!mark)
358 return true;
359
360 if (refcount_inc_not_zero(&mark->refcnt)) {
361 spin_lock(&mark->lock);
362 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) {
363 /* mark is attached, group is still alive then */
364 atomic_inc(&mark->group->user_waits);
365 spin_unlock(&mark->lock);
366 return true;
367 }
368 spin_unlock(&mark->lock);
369 fsnotify_put_mark(mark);
370 }
371 return false;
372}
373
374/*
375 * Puts marks and wakes up group destruction if necessary.
376 *
377 * Pairs with fsnotify_get_mark_safe()
378 */
379static void fsnotify_put_mark_wake(struct fsnotify_mark *mark)
380{
381 if (mark) {
382 struct fsnotify_group *group = mark->group;
383
384 fsnotify_put_mark(mark);
385 /*
386 * We abuse notification_waitq on group shutdown for waiting for
387 * all marks pinned when waiting for userspace.
388 */
389 if (atomic_dec_and_test(&group->user_waits) && group->shutdown)
390 wake_up(&group->notification_waitq);
391 }
392}
393
394bool fsnotify_prepare_user_wait(struct fsnotify_iter_info *iter_info)
395 __releases(&fsnotify_mark_srcu)
396{
397 int type;
398
399 fsnotify_foreach_iter_type(type) {
400 /* This can fail if mark is being removed */
401 if (!fsnotify_get_mark_safe(iter_info->marks[type])) {
402 __release(&fsnotify_mark_srcu);
403 goto fail;
404 }
405 }
406
407 /*
408 * Now that both marks are pinned by refcount in the inode / vfsmount
409 * lists, we can drop SRCU lock, and safely resume the list iteration
410 * once userspace returns.
411 */
412 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
413
414 return true;
415
416fail:
417 for (type--; type >= 0; type--)
418 fsnotify_put_mark_wake(iter_info->marks[type]);
419 return false;
420}
421
422void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
423 __acquires(&fsnotify_mark_srcu)
424{
425 int type;
426
427 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
428 fsnotify_foreach_iter_type(type)
429 fsnotify_put_mark_wake(iter_info->marks[type]);
430}
431
432/*
433 * Mark mark as detached, remove it from group list. Mark still stays in object
434 * list until its last reference is dropped. Note that we rely on mark being
435 * removed from group list before corresponding reference to it is dropped. In
436 * particular we rely on mark->connector being valid while we hold
437 * group->mark_mutex if we found the mark through g_list.
438 *
439 * Must be called with group->mark_mutex held. The caller must either hold
440 * reference to the mark or be protected by fsnotify_mark_srcu.
441 */
442void fsnotify_detach_mark(struct fsnotify_mark *mark)
443{
444 fsnotify_group_assert_locked(mark->group);
445 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
446 refcount_read(&mark->refcnt) < 1 +
447 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
448
449 spin_lock(&mark->lock);
450 /* something else already called this function on this mark */
451 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
452 spin_unlock(&mark->lock);
453 return;
454 }
455 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
456 list_del_init(&mark->g_list);
457 spin_unlock(&mark->lock);
458
459 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
460 fsnotify_put_mark(mark);
461}
462
463/*
464 * Free fsnotify mark. The mark is actually only marked as being freed. The
465 * freeing is actually happening only once last reference to the mark is
466 * dropped from a workqueue which first waits for srcu period end.
467 *
468 * Caller must have a reference to the mark or be protected by
469 * fsnotify_mark_srcu.
470 */
471void fsnotify_free_mark(struct fsnotify_mark *mark)
472{
473 struct fsnotify_group *group = mark->group;
474
475 spin_lock(&mark->lock);
476 /* something else already called this function on this mark */
477 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
478 spin_unlock(&mark->lock);
479 return;
480 }
481 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
482 spin_unlock(&mark->lock);
483
484 /*
485 * Some groups like to know that marks are being freed. This is a
486 * callback to the group function to let it know that this mark
487 * is being freed.
488 */
489 if (group->ops->freeing_mark)
490 group->ops->freeing_mark(mark, group);
491}
492
493void fsnotify_destroy_mark(struct fsnotify_mark *mark,
494 struct fsnotify_group *group)
495{
496 fsnotify_group_lock(group);
497 fsnotify_detach_mark(mark);
498 fsnotify_group_unlock(group);
499 fsnotify_free_mark(mark);
500}
501EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
502
503/*
504 * Sorting function for lists of fsnotify marks.
505 *
506 * Fanotify supports different notification classes (reflected as priority of
507 * notification group). Events shall be passed to notification groups in
508 * decreasing priority order. To achieve this marks in notification lists for
509 * inodes and vfsmounts are sorted so that priorities of corresponding groups
510 * are descending.
511 *
512 * Furthermore correct handling of the ignore mask requires processing inode
513 * and vfsmount marks of each group together. Using the group address as
514 * further sort criterion provides a unique sorting order and thus we can
515 * merge inode and vfsmount lists of marks in linear time and find groups
516 * present in both lists.
517 *
518 * A return value of 1 signifies that b has priority over a.
519 * A return value of 0 signifies that the two marks have to be handled together.
520 * A return value of -1 signifies that a has priority over b.
521 */
522int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
523{
524 if (a == b)
525 return 0;
526 if (!a)
527 return 1;
528 if (!b)
529 return -1;
530 if (a->priority < b->priority)
531 return 1;
532 if (a->priority > b->priority)
533 return -1;
534 if (a < b)
535 return 1;
536 return -1;
537}
538
539static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
540 unsigned int obj_type)
541{
542 struct fsnotify_mark_connector *conn;
543
544 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
545 if (!conn)
546 return -ENOMEM;
547 spin_lock_init(&conn->lock);
548 INIT_HLIST_HEAD(&conn->list);
549 conn->flags = 0;
550 conn->type = obj_type;
551 conn->obj = connp;
552 conn->flags = 0;
553 fsnotify_get_sb_connectors(conn);
554
555 /*
556 * cmpxchg() provides the barrier so that readers of *connp can see
557 * only initialized structure
558 */
559 if (cmpxchg(connp, NULL, conn)) {
560 /* Someone else created list structure for us */
561 fsnotify_put_sb_connectors(conn);
562 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
563 }
564
565 return 0;
566}
567
568/*
569 * Get mark connector, make sure it is alive and return with its lock held.
570 * This is for users that get connector pointer from inode or mount. Users that
571 * hold reference to a mark on the list may directly lock connector->lock as
572 * they are sure list cannot go away under them.
573 */
574static struct fsnotify_mark_connector *fsnotify_grab_connector(
575 fsnotify_connp_t *connp)
576{
577 struct fsnotify_mark_connector *conn;
578 int idx;
579
580 idx = srcu_read_lock(&fsnotify_mark_srcu);
581 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
582 if (!conn)
583 goto out;
584 spin_lock(&conn->lock);
585 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
586 spin_unlock(&conn->lock);
587 srcu_read_unlock(&fsnotify_mark_srcu, idx);
588 return NULL;
589 }
590out:
591 srcu_read_unlock(&fsnotify_mark_srcu, idx);
592 return conn;
593}
594
595/*
596 * Add mark into proper place in given list of marks. These marks may be used
597 * for the fsnotify backend to determine which event types should be delivered
598 * to which group and for which inodes. These marks are ordered according to
599 * priority, highest number first, and then by the group's location in memory.
600 */
601static int fsnotify_add_mark_list(struct fsnotify_mark *mark,
602 fsnotify_connp_t *connp,
603 unsigned int obj_type, int add_flags)
604{
605 struct fsnotify_mark *lmark, *last = NULL;
606 struct fsnotify_mark_connector *conn;
607 int cmp;
608 int err = 0;
609
610 if (WARN_ON(!fsnotify_valid_obj_type(obj_type)))
611 return -EINVAL;
612
613restart:
614 spin_lock(&mark->lock);
615 conn = fsnotify_grab_connector(connp);
616 if (!conn) {
617 spin_unlock(&mark->lock);
618 err = fsnotify_attach_connector_to_object(connp, obj_type);
619 if (err)
620 return err;
621 goto restart;
622 }
623
624 /* is mark the first mark? */
625 if (hlist_empty(&conn->list)) {
626 hlist_add_head_rcu(&mark->obj_list, &conn->list);
627 goto added;
628 }
629
630 /* should mark be in the middle of the current list? */
631 hlist_for_each_entry(lmark, &conn->list, obj_list) {
632 last = lmark;
633
634 if ((lmark->group == mark->group) &&
635 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
636 !(mark->group->flags & FSNOTIFY_GROUP_DUPS)) {
637 err = -EEXIST;
638 goto out_err;
639 }
640
641 cmp = fsnotify_compare_groups(lmark->group, mark->group);
642 if (cmp >= 0) {
643 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
644 goto added;
645 }
646 }
647
648 BUG_ON(last == NULL);
649 /* mark should be the last entry. last is the current last entry */
650 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
651added:
652 /*
653 * Since connector is attached to object using cmpxchg() we are
654 * guaranteed that connector initialization is fully visible by anyone
655 * seeing mark->connector set.
656 */
657 WRITE_ONCE(mark->connector, conn);
658out_err:
659 spin_unlock(&conn->lock);
660 spin_unlock(&mark->lock);
661 return err;
662}
663
664/*
665 * Attach an initialized mark to a given group and fs object.
666 * These marks may be used for the fsnotify backend to determine which
667 * event types should be delivered to which group.
668 */
669int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
670 fsnotify_connp_t *connp, unsigned int obj_type,
671 int add_flags)
672{
673 struct fsnotify_group *group = mark->group;
674 int ret = 0;
675
676 fsnotify_group_assert_locked(group);
677
678 /*
679 * LOCKING ORDER!!!!
680 * group->mark_mutex
681 * mark->lock
682 * mark->connector->lock
683 */
684 spin_lock(&mark->lock);
685 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
686
687 list_add(&mark->g_list, &group->marks_list);
688 fsnotify_get_mark(mark); /* for g_list */
689 spin_unlock(&mark->lock);
690
691 ret = fsnotify_add_mark_list(mark, connp, obj_type, add_flags);
692 if (ret)
693 goto err;
694
695 fsnotify_recalc_mask(mark->connector);
696
697 return ret;
698err:
699 spin_lock(&mark->lock);
700 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
701 FSNOTIFY_MARK_FLAG_ATTACHED);
702 list_del_init(&mark->g_list);
703 spin_unlock(&mark->lock);
704
705 fsnotify_put_mark(mark);
706 return ret;
707}
708
709int fsnotify_add_mark(struct fsnotify_mark *mark, fsnotify_connp_t *connp,
710 unsigned int obj_type, int add_flags)
711{
712 int ret;
713 struct fsnotify_group *group = mark->group;
714
715 fsnotify_group_lock(group);
716 ret = fsnotify_add_mark_locked(mark, connp, obj_type, add_flags);
717 fsnotify_group_unlock(group);
718 return ret;
719}
720EXPORT_SYMBOL_GPL(fsnotify_add_mark);
721
722/*
723 * Given a list of marks, find the mark associated with given group. If found
724 * take a reference to that mark and return it, else return NULL.
725 */
726struct fsnotify_mark *fsnotify_find_mark(fsnotify_connp_t *connp,
727 struct fsnotify_group *group)
728{
729 struct fsnotify_mark_connector *conn;
730 struct fsnotify_mark *mark;
731
732 conn = fsnotify_grab_connector(connp);
733 if (!conn)
734 return NULL;
735
736 hlist_for_each_entry(mark, &conn->list, obj_list) {
737 if (mark->group == group &&
738 (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
739 fsnotify_get_mark(mark);
740 spin_unlock(&conn->lock);
741 return mark;
742 }
743 }
744 spin_unlock(&conn->lock);
745 return NULL;
746}
747EXPORT_SYMBOL_GPL(fsnotify_find_mark);
748
749/* Clear any marks in a group with given type mask */
750void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
751 unsigned int obj_type)
752{
753 struct fsnotify_mark *lmark, *mark;
754 LIST_HEAD(to_free);
755 struct list_head *head = &to_free;
756
757 /* Skip selection step if we want to clear all marks. */
758 if (obj_type == FSNOTIFY_OBJ_TYPE_ANY) {
759 head = &group->marks_list;
760 goto clear;
761 }
762 /*
763 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
764 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
765 * to_free list so we have to use mark_mutex even when accessing that
766 * list. And freeing mark requires us to drop mark_mutex. So we can
767 * reliably free only the first mark in the list. That's why we first
768 * move marks to free to to_free list in one go and then free marks in
769 * to_free list one by one.
770 */
771 fsnotify_group_lock(group);
772 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
773 if (mark->connector->type == obj_type)
774 list_move(&mark->g_list, &to_free);
775 }
776 fsnotify_group_unlock(group);
777
778clear:
779 while (1) {
780 fsnotify_group_lock(group);
781 if (list_empty(head)) {
782 fsnotify_group_unlock(group);
783 break;
784 }
785 mark = list_first_entry(head, struct fsnotify_mark, g_list);
786 fsnotify_get_mark(mark);
787 fsnotify_detach_mark(mark);
788 fsnotify_group_unlock(group);
789 fsnotify_free_mark(mark);
790 fsnotify_put_mark(mark);
791 }
792}
793
794/* Destroy all marks attached to an object via connector */
795void fsnotify_destroy_marks(fsnotify_connp_t *connp)
796{
797 struct fsnotify_mark_connector *conn;
798 struct fsnotify_mark *mark, *old_mark = NULL;
799 void *objp;
800 unsigned int type;
801
802 conn = fsnotify_grab_connector(connp);
803 if (!conn)
804 return;
805 /*
806 * We have to be careful since we can race with e.g.
807 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
808 * list can get modified. However we are holding mark reference and
809 * thus our mark cannot be removed from obj_list so we can continue
810 * iteration after regaining conn->lock.
811 */
812 hlist_for_each_entry(mark, &conn->list, obj_list) {
813 fsnotify_get_mark(mark);
814 spin_unlock(&conn->lock);
815 if (old_mark)
816 fsnotify_put_mark(old_mark);
817 old_mark = mark;
818 fsnotify_destroy_mark(mark, mark->group);
819 spin_lock(&conn->lock);
820 }
821 /*
822 * Detach list from object now so that we don't pin inode until all
823 * mark references get dropped. It would lead to strange results such
824 * as delaying inode deletion or blocking unmount.
825 */
826 objp = fsnotify_detach_connector_from_object(conn, &type);
827 spin_unlock(&conn->lock);
828 if (old_mark)
829 fsnotify_put_mark(old_mark);
830 fsnotify_drop_object(type, objp);
831}
832
833/*
834 * Nothing fancy, just initialize lists and locks and counters.
835 */
836void fsnotify_init_mark(struct fsnotify_mark *mark,
837 struct fsnotify_group *group)
838{
839 memset(mark, 0, sizeof(*mark));
840 spin_lock_init(&mark->lock);
841 refcount_set(&mark->refcnt, 1);
842 fsnotify_get_group(group);
843 mark->group = group;
844 WRITE_ONCE(mark->connector, NULL);
845}
846EXPORT_SYMBOL_GPL(fsnotify_init_mark);
847
848/*
849 * Destroy all marks in destroy_list, waits for SRCU period to finish before
850 * actually freeing marks.
851 */
852static void fsnotify_mark_destroy_workfn(struct work_struct *work)
853{
854 struct fsnotify_mark *mark, *next;
855 struct list_head private_destroy_list;
856
857 spin_lock(&destroy_lock);
858 /* exchange the list head */
859 list_replace_init(&destroy_list, &private_destroy_list);
860 spin_unlock(&destroy_lock);
861
862 synchronize_srcu(&fsnotify_mark_srcu);
863
864 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
865 list_del_init(&mark->g_list);
866 fsnotify_final_mark_destroy(mark);
867 }
868}
869
870/* Wait for all marks queued for destruction to be actually destroyed */
871void fsnotify_wait_marks_destroyed(void)
872{
873 flush_delayed_work(&reaper_work);
874}
875EXPORT_SYMBOL_GPL(fsnotify_wait_marks_destroyed);