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