Loading...
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 __releases(&fsnotify_mark_srcu)
329{
330 int type;
331
332 fsnotify_foreach_obj_type(type) {
333 /* This can fail if mark is being removed */
334 if (!fsnotify_get_mark_safe(iter_info->marks[type])) {
335 __release(&fsnotify_mark_srcu);
336 goto fail;
337 }
338 }
339
340 /*
341 * Now that both marks are pinned by refcount in the inode / vfsmount
342 * lists, we can drop SRCU lock, and safely resume the list iteration
343 * once userspace returns.
344 */
345 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
346
347 return true;
348
349fail:
350 for (type--; type >= 0; type--)
351 fsnotify_put_mark_wake(iter_info->marks[type]);
352 return false;
353}
354
355void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
356 __acquires(&fsnotify_mark_srcu)
357{
358 int type;
359
360 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
361 fsnotify_foreach_obj_type(type)
362 fsnotify_put_mark_wake(iter_info->marks[type]);
363}
364
365/*
366 * Mark mark as detached, remove it from group list. Mark still stays in object
367 * list until its last reference is dropped. Note that we rely on mark being
368 * removed from group list before corresponding reference to it is dropped. In
369 * particular we rely on mark->connector being valid while we hold
370 * group->mark_mutex if we found the mark through g_list.
371 *
372 * Must be called with group->mark_mutex held. The caller must either hold
373 * reference to the mark or be protected by fsnotify_mark_srcu.
374 */
375void fsnotify_detach_mark(struct fsnotify_mark *mark)
376{
377 struct fsnotify_group *group = mark->group;
378
379 WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
380 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
381 refcount_read(&mark->refcnt) < 1 +
382 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
383
384 spin_lock(&mark->lock);
385 /* something else already called this function on this mark */
386 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
387 spin_unlock(&mark->lock);
388 return;
389 }
390 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
391 list_del_init(&mark->g_list);
392 spin_unlock(&mark->lock);
393
394 atomic_dec(&group->num_marks);
395
396 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
397 fsnotify_put_mark(mark);
398}
399
400/*
401 * Free fsnotify mark. The mark is actually only marked as being freed. The
402 * freeing is actually happening only once last reference to the mark is
403 * dropped from a workqueue which first waits for srcu period end.
404 *
405 * Caller must have a reference to the mark or be protected by
406 * fsnotify_mark_srcu.
407 */
408void fsnotify_free_mark(struct fsnotify_mark *mark)
409{
410 struct fsnotify_group *group = mark->group;
411
412 spin_lock(&mark->lock);
413 /* something else already called this function on this mark */
414 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
415 spin_unlock(&mark->lock);
416 return;
417 }
418 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
419 spin_unlock(&mark->lock);
420
421 /*
422 * Some groups like to know that marks are being freed. This is a
423 * callback to the group function to let it know that this mark
424 * is being freed.
425 */
426 if (group->ops->freeing_mark)
427 group->ops->freeing_mark(mark, group);
428}
429
430void fsnotify_destroy_mark(struct fsnotify_mark *mark,
431 struct fsnotify_group *group)
432{
433 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
434 fsnotify_detach_mark(mark);
435 mutex_unlock(&group->mark_mutex);
436 fsnotify_free_mark(mark);
437}
438EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
439
440/*
441 * Sorting function for lists of fsnotify marks.
442 *
443 * Fanotify supports different notification classes (reflected as priority of
444 * notification group). Events shall be passed to notification groups in
445 * decreasing priority order. To achieve this marks in notification lists for
446 * inodes and vfsmounts are sorted so that priorities of corresponding groups
447 * are descending.
448 *
449 * Furthermore correct handling of the ignore mask requires processing inode
450 * and vfsmount marks of each group together. Using the group address as
451 * further sort criterion provides a unique sorting order and thus we can
452 * merge inode and vfsmount lists of marks in linear time and find groups
453 * present in both lists.
454 *
455 * A return value of 1 signifies that b has priority over a.
456 * A return value of 0 signifies that the two marks have to be handled together.
457 * A return value of -1 signifies that a has priority over b.
458 */
459int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
460{
461 if (a == b)
462 return 0;
463 if (!a)
464 return 1;
465 if (!b)
466 return -1;
467 if (a->priority < b->priority)
468 return 1;
469 if (a->priority > b->priority)
470 return -1;
471 if (a < b)
472 return 1;
473 return -1;
474}
475
476static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
477 unsigned int type,
478 __kernel_fsid_t *fsid)
479{
480 struct inode *inode = NULL;
481 struct fsnotify_mark_connector *conn;
482
483 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
484 if (!conn)
485 return -ENOMEM;
486 spin_lock_init(&conn->lock);
487 INIT_HLIST_HEAD(&conn->list);
488 conn->type = type;
489 conn->obj = connp;
490 /* Cache fsid of filesystem containing the object */
491 if (fsid) {
492 conn->fsid = *fsid;
493 conn->flags = FSNOTIFY_CONN_FLAG_HAS_FSID;
494 } else {
495 conn->fsid.val[0] = conn->fsid.val[1] = 0;
496 conn->flags = 0;
497 }
498 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
499 inode = igrab(fsnotify_conn_inode(conn));
500 /*
501 * cmpxchg() provides the barrier so that readers of *connp can see
502 * only initialized structure
503 */
504 if (cmpxchg(connp, NULL, conn)) {
505 /* Someone else created list structure for us */
506 if (inode)
507 iput(inode);
508 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
509 }
510
511 return 0;
512}
513
514/*
515 * Get mark connector, make sure it is alive and return with its lock held.
516 * This is for users that get connector pointer from inode or mount. Users that
517 * hold reference to a mark on the list may directly lock connector->lock as
518 * they are sure list cannot go away under them.
519 */
520static struct fsnotify_mark_connector *fsnotify_grab_connector(
521 fsnotify_connp_t *connp)
522{
523 struct fsnotify_mark_connector *conn;
524 int idx;
525
526 idx = srcu_read_lock(&fsnotify_mark_srcu);
527 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
528 if (!conn)
529 goto out;
530 spin_lock(&conn->lock);
531 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
532 spin_unlock(&conn->lock);
533 srcu_read_unlock(&fsnotify_mark_srcu, idx);
534 return NULL;
535 }
536out:
537 srcu_read_unlock(&fsnotify_mark_srcu, idx);
538 return conn;
539}
540
541/*
542 * Add mark into proper place in given list of marks. These marks may be used
543 * for the fsnotify backend to determine which event types should be delivered
544 * to which group and for which inodes. These marks are ordered according to
545 * priority, highest number first, and then by the group's location in memory.
546 */
547static int fsnotify_add_mark_list(struct fsnotify_mark *mark,
548 fsnotify_connp_t *connp, unsigned int type,
549 int allow_dups, __kernel_fsid_t *fsid)
550{
551 struct fsnotify_mark *lmark, *last = NULL;
552 struct fsnotify_mark_connector *conn;
553 int cmp;
554 int err = 0;
555
556 if (WARN_ON(!fsnotify_valid_obj_type(type)))
557 return -EINVAL;
558
559 /* Backend is expected to check for zero fsid (e.g. tmpfs) */
560 if (fsid && WARN_ON_ONCE(!fsid->val[0] && !fsid->val[1]))
561 return -ENODEV;
562
563restart:
564 spin_lock(&mark->lock);
565 conn = fsnotify_grab_connector(connp);
566 if (!conn) {
567 spin_unlock(&mark->lock);
568 err = fsnotify_attach_connector_to_object(connp, type, fsid);
569 if (err)
570 return err;
571 goto restart;
572 } else if (fsid && !(conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID)) {
573 conn->fsid = *fsid;
574 /* Pairs with smp_rmb() in fanotify_get_fsid() */
575 smp_wmb();
576 conn->flags |= FSNOTIFY_CONN_FLAG_HAS_FSID;
577 } else if (fsid && (conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID) &&
578 (fsid->val[0] != conn->fsid.val[0] ||
579 fsid->val[1] != conn->fsid.val[1])) {
580 /*
581 * Backend is expected to check for non uniform fsid
582 * (e.g. btrfs), but maybe we missed something?
583 * Only allow setting conn->fsid once to non zero fsid.
584 * inotify and non-fid fanotify groups do not set nor test
585 * conn->fsid.
586 */
587 pr_warn_ratelimited("%s: fsid mismatch on object of type %u: "
588 "%x.%x != %x.%x\n", __func__, conn->type,
589 fsid->val[0], fsid->val[1],
590 conn->fsid.val[0], conn->fsid.val[1]);
591 err = -EXDEV;
592 goto out_err;
593 }
594
595 /* is mark the first mark? */
596 if (hlist_empty(&conn->list)) {
597 hlist_add_head_rcu(&mark->obj_list, &conn->list);
598 goto added;
599 }
600
601 /* should mark be in the middle of the current list? */
602 hlist_for_each_entry(lmark, &conn->list, obj_list) {
603 last = lmark;
604
605 if ((lmark->group == mark->group) &&
606 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
607 !allow_dups) {
608 err = -EEXIST;
609 goto out_err;
610 }
611
612 cmp = fsnotify_compare_groups(lmark->group, mark->group);
613 if (cmp >= 0) {
614 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
615 goto added;
616 }
617 }
618
619 BUG_ON(last == NULL);
620 /* mark should be the last entry. last is the current last entry */
621 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
622added:
623 /*
624 * Since connector is attached to object using cmpxchg() we are
625 * guaranteed that connector initialization is fully visible by anyone
626 * seeing mark->connector set.
627 */
628 WRITE_ONCE(mark->connector, conn);
629out_err:
630 spin_unlock(&conn->lock);
631 spin_unlock(&mark->lock);
632 return err;
633}
634
635/*
636 * Attach an initialized mark to a given group and fs object.
637 * These marks may be used for the fsnotify backend to determine which
638 * event types should be delivered to which group.
639 */
640int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
641 fsnotify_connp_t *connp, unsigned int type,
642 int allow_dups, __kernel_fsid_t *fsid)
643{
644 struct fsnotify_group *group = mark->group;
645 int ret = 0;
646
647 BUG_ON(!mutex_is_locked(&group->mark_mutex));
648
649 /*
650 * LOCKING ORDER!!!!
651 * group->mark_mutex
652 * mark->lock
653 * mark->connector->lock
654 */
655 spin_lock(&mark->lock);
656 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
657
658 list_add(&mark->g_list, &group->marks_list);
659 atomic_inc(&group->num_marks);
660 fsnotify_get_mark(mark); /* for g_list */
661 spin_unlock(&mark->lock);
662
663 ret = fsnotify_add_mark_list(mark, connp, type, allow_dups, fsid);
664 if (ret)
665 goto err;
666
667 if (mark->mask)
668 fsnotify_recalc_mask(mark->connector);
669
670 return ret;
671err:
672 spin_lock(&mark->lock);
673 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
674 FSNOTIFY_MARK_FLAG_ATTACHED);
675 list_del_init(&mark->g_list);
676 spin_unlock(&mark->lock);
677 atomic_dec(&group->num_marks);
678
679 fsnotify_put_mark(mark);
680 return ret;
681}
682
683int fsnotify_add_mark(struct fsnotify_mark *mark, fsnotify_connp_t *connp,
684 unsigned int type, int allow_dups, __kernel_fsid_t *fsid)
685{
686 int ret;
687 struct fsnotify_group *group = mark->group;
688
689 mutex_lock(&group->mark_mutex);
690 ret = fsnotify_add_mark_locked(mark, connp, type, allow_dups, fsid);
691 mutex_unlock(&group->mark_mutex);
692 return ret;
693}
694EXPORT_SYMBOL_GPL(fsnotify_add_mark);
695
696/*
697 * Given a list of marks, find the mark associated with given group. If found
698 * take a reference to that mark and return it, else return NULL.
699 */
700struct fsnotify_mark *fsnotify_find_mark(fsnotify_connp_t *connp,
701 struct fsnotify_group *group)
702{
703 struct fsnotify_mark_connector *conn;
704 struct fsnotify_mark *mark;
705
706 conn = fsnotify_grab_connector(connp);
707 if (!conn)
708 return NULL;
709
710 hlist_for_each_entry(mark, &conn->list, obj_list) {
711 if (mark->group == group &&
712 (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
713 fsnotify_get_mark(mark);
714 spin_unlock(&conn->lock);
715 return mark;
716 }
717 }
718 spin_unlock(&conn->lock);
719 return NULL;
720}
721EXPORT_SYMBOL_GPL(fsnotify_find_mark);
722
723/* Clear any marks in a group with given type mask */
724void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
725 unsigned int type_mask)
726{
727 struct fsnotify_mark *lmark, *mark;
728 LIST_HEAD(to_free);
729 struct list_head *head = &to_free;
730
731 /* Skip selection step if we want to clear all marks. */
732 if (type_mask == FSNOTIFY_OBJ_ALL_TYPES_MASK) {
733 head = &group->marks_list;
734 goto clear;
735 }
736 /*
737 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
738 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
739 * to_free list so we have to use mark_mutex even when accessing that
740 * list. And freeing mark requires us to drop mark_mutex. So we can
741 * reliably free only the first mark in the list. That's why we first
742 * move marks to free to to_free list in one go and then free marks in
743 * to_free list one by one.
744 */
745 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
746 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
747 if ((1U << mark->connector->type) & type_mask)
748 list_move(&mark->g_list, &to_free);
749 }
750 mutex_unlock(&group->mark_mutex);
751
752clear:
753 while (1) {
754 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
755 if (list_empty(head)) {
756 mutex_unlock(&group->mark_mutex);
757 break;
758 }
759 mark = list_first_entry(head, struct fsnotify_mark, g_list);
760 fsnotify_get_mark(mark);
761 fsnotify_detach_mark(mark);
762 mutex_unlock(&group->mark_mutex);
763 fsnotify_free_mark(mark);
764 fsnotify_put_mark(mark);
765 }
766}
767
768/* Destroy all marks attached to an object via connector */
769void fsnotify_destroy_marks(fsnotify_connp_t *connp)
770{
771 struct fsnotify_mark_connector *conn;
772 struct fsnotify_mark *mark, *old_mark = NULL;
773 void *objp;
774 unsigned int type;
775
776 conn = fsnotify_grab_connector(connp);
777 if (!conn)
778 return;
779 /*
780 * We have to be careful since we can race with e.g.
781 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
782 * list can get modified. However we are holding mark reference and
783 * thus our mark cannot be removed from obj_list so we can continue
784 * iteration after regaining conn->lock.
785 */
786 hlist_for_each_entry(mark, &conn->list, obj_list) {
787 fsnotify_get_mark(mark);
788 spin_unlock(&conn->lock);
789 if (old_mark)
790 fsnotify_put_mark(old_mark);
791 old_mark = mark;
792 fsnotify_destroy_mark(mark, mark->group);
793 spin_lock(&conn->lock);
794 }
795 /*
796 * Detach list from object now so that we don't pin inode until all
797 * mark references get dropped. It would lead to strange results such
798 * as delaying inode deletion or blocking unmount.
799 */
800 objp = fsnotify_detach_connector_from_object(conn, &type);
801 spin_unlock(&conn->lock);
802 if (old_mark)
803 fsnotify_put_mark(old_mark);
804 fsnotify_drop_object(type, objp);
805}
806
807/*
808 * Nothing fancy, just initialize lists and locks and counters.
809 */
810void fsnotify_init_mark(struct fsnotify_mark *mark,
811 struct fsnotify_group *group)
812{
813 memset(mark, 0, sizeof(*mark));
814 spin_lock_init(&mark->lock);
815 refcount_set(&mark->refcnt, 1);
816 fsnotify_get_group(group);
817 mark->group = group;
818 WRITE_ONCE(mark->connector, NULL);
819}
820EXPORT_SYMBOL_GPL(fsnotify_init_mark);
821
822/*
823 * Destroy all marks in destroy_list, waits for SRCU period to finish before
824 * actually freeing marks.
825 */
826static void fsnotify_mark_destroy_workfn(struct work_struct *work)
827{
828 struct fsnotify_mark *mark, *next;
829 struct list_head private_destroy_list;
830
831 spin_lock(&destroy_lock);
832 /* exchange the list head */
833 list_replace_init(&destroy_list, &private_destroy_list);
834 spin_unlock(&destroy_lock);
835
836 synchronize_srcu(&fsnotify_mark_srcu);
837
838 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
839 list_del_init(&mark->g_list);
840 fsnotify_final_mark_destroy(mark);
841 }
842}
843
844/* Wait for all marks queued for destruction to be actually destroyed */
845void fsnotify_wait_marks_destroyed(void)
846{
847 flush_delayed_work(&reaper_work);
848}
849EXPORT_SYMBOL_GPL(fsnotify_wait_marks_destroyed);
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 __releases(&fsnotify_mark_srcu)
329{
330 int type;
331
332 fsnotify_foreach_obj_type(type) {
333 /* This can fail if mark is being removed */
334 if (!fsnotify_get_mark_safe(iter_info->marks[type])) {
335 __release(&fsnotify_mark_srcu);
336 goto fail;
337 }
338 }
339
340 /*
341 * Now that both marks are pinned by refcount in the inode / vfsmount
342 * lists, we can drop SRCU lock, and safely resume the list iteration
343 * once userspace returns.
344 */
345 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
346
347 return true;
348
349fail:
350 for (type--; type >= 0; type--)
351 fsnotify_put_mark_wake(iter_info->marks[type]);
352 return false;
353}
354
355void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
356 __acquires(&fsnotify_mark_srcu)
357{
358 int type;
359
360 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
361 fsnotify_foreach_obj_type(type)
362 fsnotify_put_mark_wake(iter_info->marks[type]);
363}
364
365/*
366 * Mark mark as detached, remove it from group list. Mark still stays in object
367 * list until its last reference is dropped. Note that we rely on mark being
368 * removed from group list before corresponding reference to it is dropped. In
369 * particular we rely on mark->connector being valid while we hold
370 * group->mark_mutex if we found the mark through g_list.
371 *
372 * Must be called with group->mark_mutex held. The caller must either hold
373 * reference to the mark or be protected by fsnotify_mark_srcu.
374 */
375void fsnotify_detach_mark(struct fsnotify_mark *mark)
376{
377 struct fsnotify_group *group = mark->group;
378
379 WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
380 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
381 refcount_read(&mark->refcnt) < 1 +
382 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
383
384 spin_lock(&mark->lock);
385 /* something else already called this function on this mark */
386 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
387 spin_unlock(&mark->lock);
388 return;
389 }
390 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
391 list_del_init(&mark->g_list);
392 spin_unlock(&mark->lock);
393
394 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
395 fsnotify_put_mark(mark);
396}
397
398/*
399 * Free fsnotify mark. The mark is actually only marked as being freed. The
400 * freeing is actually happening only once last reference to the mark is
401 * dropped from a workqueue which first waits for srcu period end.
402 *
403 * Caller must have a reference to the mark or be protected by
404 * fsnotify_mark_srcu.
405 */
406void fsnotify_free_mark(struct fsnotify_mark *mark)
407{
408 struct fsnotify_group *group = mark->group;
409
410 spin_lock(&mark->lock);
411 /* something else already called this function on this mark */
412 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
413 spin_unlock(&mark->lock);
414 return;
415 }
416 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
417 spin_unlock(&mark->lock);
418
419 /*
420 * Some groups like to know that marks are being freed. This is a
421 * callback to the group function to let it know that this mark
422 * is being freed.
423 */
424 if (group->ops->freeing_mark)
425 group->ops->freeing_mark(mark, group);
426}
427
428void fsnotify_destroy_mark(struct fsnotify_mark *mark,
429 struct fsnotify_group *group)
430{
431 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
432 fsnotify_detach_mark(mark);
433 mutex_unlock(&group->mark_mutex);
434 fsnotify_free_mark(mark);
435}
436EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
437
438/*
439 * Sorting function for lists of fsnotify marks.
440 *
441 * Fanotify supports different notification classes (reflected as priority of
442 * notification group). Events shall be passed to notification groups in
443 * decreasing priority order. To achieve this marks in notification lists for
444 * inodes and vfsmounts are sorted so that priorities of corresponding groups
445 * are descending.
446 *
447 * Furthermore correct handling of the ignore mask requires processing inode
448 * and vfsmount marks of each group together. Using the group address as
449 * further sort criterion provides a unique sorting order and thus we can
450 * merge inode and vfsmount lists of marks in linear time and find groups
451 * present in both lists.
452 *
453 * A return value of 1 signifies that b has priority over a.
454 * A return value of 0 signifies that the two marks have to be handled together.
455 * A return value of -1 signifies that a has priority over b.
456 */
457int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
458{
459 if (a == b)
460 return 0;
461 if (!a)
462 return 1;
463 if (!b)
464 return -1;
465 if (a->priority < b->priority)
466 return 1;
467 if (a->priority > b->priority)
468 return -1;
469 if (a < b)
470 return 1;
471 return -1;
472}
473
474static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
475 unsigned int type,
476 __kernel_fsid_t *fsid)
477{
478 struct inode *inode = NULL;
479 struct fsnotify_mark_connector *conn;
480
481 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
482 if (!conn)
483 return -ENOMEM;
484 spin_lock_init(&conn->lock);
485 INIT_HLIST_HEAD(&conn->list);
486 conn->type = type;
487 conn->obj = connp;
488 /* Cache fsid of filesystem containing the object */
489 if (fsid) {
490 conn->fsid = *fsid;
491 conn->flags = FSNOTIFY_CONN_FLAG_HAS_FSID;
492 } else {
493 conn->fsid.val[0] = conn->fsid.val[1] = 0;
494 conn->flags = 0;
495 }
496 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
497 inode = igrab(fsnotify_conn_inode(conn));
498 /*
499 * cmpxchg() provides the barrier so that readers of *connp can see
500 * only initialized structure
501 */
502 if (cmpxchg(connp, NULL, conn)) {
503 /* Someone else created list structure for us */
504 if (inode)
505 iput(inode);
506 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
507 }
508
509 return 0;
510}
511
512/*
513 * Get mark connector, make sure it is alive and return with its lock held.
514 * This is for users that get connector pointer from inode or mount. Users that
515 * hold reference to a mark on the list may directly lock connector->lock as
516 * they are sure list cannot go away under them.
517 */
518static struct fsnotify_mark_connector *fsnotify_grab_connector(
519 fsnotify_connp_t *connp)
520{
521 struct fsnotify_mark_connector *conn;
522 int idx;
523
524 idx = srcu_read_lock(&fsnotify_mark_srcu);
525 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
526 if (!conn)
527 goto out;
528 spin_lock(&conn->lock);
529 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
530 spin_unlock(&conn->lock);
531 srcu_read_unlock(&fsnotify_mark_srcu, idx);
532 return NULL;
533 }
534out:
535 srcu_read_unlock(&fsnotify_mark_srcu, idx);
536 return conn;
537}
538
539/*
540 * Add mark into proper place in given list of marks. These marks may be used
541 * for the fsnotify backend to determine which event types should be delivered
542 * to which group and for which inodes. These marks are ordered according to
543 * priority, highest number first, and then by the group's location in memory.
544 */
545static int fsnotify_add_mark_list(struct fsnotify_mark *mark,
546 fsnotify_connp_t *connp, unsigned int type,
547 int allow_dups, __kernel_fsid_t *fsid)
548{
549 struct fsnotify_mark *lmark, *last = NULL;
550 struct fsnotify_mark_connector *conn;
551 int cmp;
552 int err = 0;
553
554 if (WARN_ON(!fsnotify_valid_obj_type(type)))
555 return -EINVAL;
556
557 /* Backend is expected to check for zero fsid (e.g. tmpfs) */
558 if (fsid && WARN_ON_ONCE(!fsid->val[0] && !fsid->val[1]))
559 return -ENODEV;
560
561restart:
562 spin_lock(&mark->lock);
563 conn = fsnotify_grab_connector(connp);
564 if (!conn) {
565 spin_unlock(&mark->lock);
566 err = fsnotify_attach_connector_to_object(connp, type, fsid);
567 if (err)
568 return err;
569 goto restart;
570 } else if (fsid && !(conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID)) {
571 conn->fsid = *fsid;
572 /* Pairs with smp_rmb() in fanotify_get_fsid() */
573 smp_wmb();
574 conn->flags |= FSNOTIFY_CONN_FLAG_HAS_FSID;
575 } else if (fsid && (conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID) &&
576 (fsid->val[0] != conn->fsid.val[0] ||
577 fsid->val[1] != conn->fsid.val[1])) {
578 /*
579 * Backend is expected to check for non uniform fsid
580 * (e.g. btrfs), but maybe we missed something?
581 * Only allow setting conn->fsid once to non zero fsid.
582 * inotify and non-fid fanotify groups do not set nor test
583 * conn->fsid.
584 */
585 pr_warn_ratelimited("%s: fsid mismatch on object of type %u: "
586 "%x.%x != %x.%x\n", __func__, conn->type,
587 fsid->val[0], fsid->val[1],
588 conn->fsid.val[0], conn->fsid.val[1]);
589 err = -EXDEV;
590 goto out_err;
591 }
592
593 /* is mark the first mark? */
594 if (hlist_empty(&conn->list)) {
595 hlist_add_head_rcu(&mark->obj_list, &conn->list);
596 goto added;
597 }
598
599 /* should mark be in the middle of the current list? */
600 hlist_for_each_entry(lmark, &conn->list, obj_list) {
601 last = lmark;
602
603 if ((lmark->group == mark->group) &&
604 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
605 !allow_dups) {
606 err = -EEXIST;
607 goto out_err;
608 }
609
610 cmp = fsnotify_compare_groups(lmark->group, mark->group);
611 if (cmp >= 0) {
612 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
613 goto added;
614 }
615 }
616
617 BUG_ON(last == NULL);
618 /* mark should be the last entry. last is the current last entry */
619 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
620added:
621 /*
622 * Since connector is attached to object using cmpxchg() we are
623 * guaranteed that connector initialization is fully visible by anyone
624 * seeing mark->connector set.
625 */
626 WRITE_ONCE(mark->connector, conn);
627out_err:
628 spin_unlock(&conn->lock);
629 spin_unlock(&mark->lock);
630 return err;
631}
632
633/*
634 * Attach an initialized mark to a given group and fs object.
635 * These marks may be used for the fsnotify backend to determine which
636 * event types should be delivered to which group.
637 */
638int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
639 fsnotify_connp_t *connp, unsigned int type,
640 int allow_dups, __kernel_fsid_t *fsid)
641{
642 struct fsnotify_group *group = mark->group;
643 int ret = 0;
644
645 BUG_ON(!mutex_is_locked(&group->mark_mutex));
646
647 /*
648 * LOCKING ORDER!!!!
649 * group->mark_mutex
650 * mark->lock
651 * mark->connector->lock
652 */
653 spin_lock(&mark->lock);
654 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
655
656 list_add(&mark->g_list, &group->marks_list);
657 fsnotify_get_mark(mark); /* for g_list */
658 spin_unlock(&mark->lock);
659
660 ret = fsnotify_add_mark_list(mark, connp, type, allow_dups, fsid);
661 if (ret)
662 goto err;
663
664 if (mark->mask)
665 fsnotify_recalc_mask(mark->connector);
666
667 return ret;
668err:
669 spin_lock(&mark->lock);
670 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
671 FSNOTIFY_MARK_FLAG_ATTACHED);
672 list_del_init(&mark->g_list);
673 spin_unlock(&mark->lock);
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);