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 mark->refcnt tells how many "things" in the kernel currently are
24 * referencing this object. The object typically will live inside the kernel
25 * with a refcnt of 2, one for each list it is on (i_list, g_list). Any task
26 * which can find this object holding the appropriete locks, can take a reference
27 * and the object itself is guaranteed to survive until the reference is dropped.
28 *
29 * LOCKING:
30 * There are 3 spinlocks involved with fsnotify inode marks and they MUST
31 * be taken in order as follows:
32 *
33 * mark->lock
34 * group->mark_lock
35 * inode->i_lock
36 *
37 * mark->lock protects 2 things, mark->group and mark->inode. You must hold
38 * that lock to dereference either of these things (they could be NULL even with
39 * the lock)
40 *
41 * group->mark_lock protects the marks_list anchored inside a given group
42 * and each mark is hooked via the g_list. It also sorta protects the
43 * free_g_list, which when used is anchored by a private list on the stack of the
44 * task which held the group->mark_lock.
45 *
46 * inode->i_lock protects the i_fsnotify_marks list anchored inside a
47 * given inode and each mark is hooked via the i_list. (and sorta the
48 * free_i_list)
49 *
50 *
51 * LIFETIME:
52 * Inode marks survive between when they are added to an inode and when their
53 * refcnt==0.
54 *
55 * The inode mark can be cleared for a number of different reasons including:
56 * - The inode is unlinked for the last time. (fsnotify_inode_remove)
57 * - The inode is being evicted from cache. (fsnotify_inode_delete)
58 * - The fs the inode is on is unmounted. (fsnotify_inode_delete/fsnotify_unmount_inodes)
59 * - Something explicitly requests that it be removed. (fsnotify_destroy_mark)
60 * - The fsnotify_group associated with the mark is going away and all such marks
61 * need to be cleaned up. (fsnotify_clear_marks_by_group)
62 *
63 * Worst case we are given an inode and need to clean up all the marks on that
64 * inode. We take i_lock and walk the i_fsnotify_marks safely. For each
65 * mark on the list we take a reference (so the mark can't disappear under us).
66 * We remove that mark form the inode's list of marks and we add this mark to a
67 * private list anchored on the stack using i_free_list; At this point we no
68 * longer fear anything finding the mark using the inode's list of marks.
69 *
70 * We can safely and locklessly run the private list on the stack of everything
71 * we just unattached from the original inode. For each mark on the private list
72 * we grab the mark-> and can thus dereference mark->group and mark->inode. If
73 * we see the group and inode are not NULL we take those locks. Now holding all
74 * 3 locks we can completely remove the mark from other tasks finding it in the
75 * future. Remember, 10 things might already be referencing this mark, but they
76 * better be holding a ref. We drop our reference we took before we unhooked it
77 * from the inode. When the ref hits 0 we can free the mark.
78 *
79 * Very similarly for freeing by group, except we use free_g_list.
80 *
81 * This has the very interesting property of being able to run concurrently with
82 * any (or all) other directions.
83 */
84
85#include <linux/fs.h>
86#include <linux/init.h>
87#include <linux/kernel.h>
88#include <linux/kthread.h>
89#include <linux/module.h>
90#include <linux/mutex.h>
91#include <linux/slab.h>
92#include <linux/spinlock.h>
93#include <linux/srcu.h>
94
95#include <linux/atomic.h>
96
97#include <linux/fsnotify_backend.h>
98#include "fsnotify.h"
99
100struct srcu_struct fsnotify_mark_srcu;
101static DEFINE_SPINLOCK(destroy_lock);
102static LIST_HEAD(destroy_list);
103static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
104
105void fsnotify_get_mark(struct fsnotify_mark *mark)
106{
107 atomic_inc(&mark->refcnt);
108}
109
110void fsnotify_put_mark(struct fsnotify_mark *mark)
111{
112 if (atomic_dec_and_test(&mark->refcnt))
113 mark->free_mark(mark);
114}
115
116/*
117 * Any time a mark is getting freed we end up here.
118 * The caller had better be holding a reference to this mark so we don't actually
119 * do the final put under the mark->lock
120 */
121void fsnotify_destroy_mark(struct fsnotify_mark *mark)
122{
123 struct fsnotify_group *group;
124 struct inode *inode = NULL;
125
126 spin_lock(&mark->lock);
127
128 group = mark->group;
129
130 /* something else already called this function on this mark */
131 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
132 spin_unlock(&mark->lock);
133 return;
134 }
135
136 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
137
138 /* 1 from caller and 1 for being on i_list/g_list */
139 BUG_ON(atomic_read(&mark->refcnt) < 2);
140
141 spin_lock(&group->mark_lock);
142
143 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
144 inode = mark->i.inode;
145 fsnotify_destroy_inode_mark(mark);
146 } else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
147 fsnotify_destroy_vfsmount_mark(mark);
148 else
149 BUG();
150
151 list_del_init(&mark->g_list);
152
153 spin_unlock(&group->mark_lock);
154 spin_unlock(&mark->lock);
155
156 spin_lock(&destroy_lock);
157 list_add(&mark->destroy_list, &destroy_list);
158 spin_unlock(&destroy_lock);
159 wake_up(&destroy_waitq);
160
161 /*
162 * Some groups like to know that marks are being freed. This is a
163 * callback to the group function to let it know that this mark
164 * is being freed.
165 */
166 if (group->ops->freeing_mark)
167 group->ops->freeing_mark(mark, group);
168
169 /*
170 * __fsnotify_update_child_dentry_flags(inode);
171 *
172 * I really want to call that, but we can't, we have no idea if the inode
173 * still exists the second we drop the mark->lock.
174 *
175 * The next time an event arrive to this inode from one of it's children
176 * __fsnotify_parent will see that the inode doesn't care about it's
177 * children and will update all of these flags then. So really this
178 * is just a lazy update (and could be a perf win...)
179 */
180
181 if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
182 iput(inode);
183
184 /*
185 * it's possible that this group tried to destroy itself, but this
186 * this mark was simultaneously being freed by inode. If that's the
187 * case, we finish freeing the group here.
188 */
189 if (unlikely(atomic_dec_and_test(&group->num_marks)))
190 fsnotify_final_destroy_group(group);
191}
192
193void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
194{
195 assert_spin_locked(&mark->lock);
196
197 mark->mask = mask;
198
199 if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
200 fsnotify_set_inode_mark_mask_locked(mark, mask);
201}
202
203void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
204{
205 assert_spin_locked(&mark->lock);
206
207 mark->ignored_mask = mask;
208}
209
210/*
211 * Attach an initialized mark to a given group and fs object.
212 * These marks may be used for the fsnotify backend to determine which
213 * event types should be delivered to which group.
214 */
215int fsnotify_add_mark(struct fsnotify_mark *mark,
216 struct fsnotify_group *group, struct inode *inode,
217 struct vfsmount *mnt, int allow_dups)
218{
219 int ret = 0;
220
221 BUG_ON(inode && mnt);
222 BUG_ON(!inode && !mnt);
223
224 /*
225 * LOCKING ORDER!!!!
226 * mark->lock
227 * group->mark_lock
228 * inode->i_lock
229 */
230 spin_lock(&mark->lock);
231 spin_lock(&group->mark_lock);
232
233 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE;
234
235 mark->group = group;
236 list_add(&mark->g_list, &group->marks_list);
237 atomic_inc(&group->num_marks);
238 fsnotify_get_mark(mark); /* for i_list and g_list */
239
240 if (inode) {
241 ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
242 if (ret)
243 goto err;
244 } else if (mnt) {
245 ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
246 if (ret)
247 goto err;
248 } else {
249 BUG();
250 }
251
252 spin_unlock(&group->mark_lock);
253
254 /* this will pin the object if appropriate */
255 fsnotify_set_mark_mask_locked(mark, mark->mask);
256
257 spin_unlock(&mark->lock);
258
259 if (inode)
260 __fsnotify_update_child_dentry_flags(inode);
261
262 return ret;
263err:
264 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
265 list_del_init(&mark->g_list);
266 mark->group = NULL;
267 atomic_dec(&group->num_marks);
268
269 spin_unlock(&group->mark_lock);
270 spin_unlock(&mark->lock);
271
272 spin_lock(&destroy_lock);
273 list_add(&mark->destroy_list, &destroy_list);
274 spin_unlock(&destroy_lock);
275 wake_up(&destroy_waitq);
276
277 return ret;
278}
279
280/*
281 * clear any marks in a group in which mark->flags & flags is true
282 */
283void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
284 unsigned int flags)
285{
286 struct fsnotify_mark *lmark, *mark;
287 LIST_HEAD(free_list);
288
289 spin_lock(&group->mark_lock);
290 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
291 if (mark->flags & flags) {
292 list_add(&mark->free_g_list, &free_list);
293 list_del_init(&mark->g_list);
294 fsnotify_get_mark(mark);
295 }
296 }
297 spin_unlock(&group->mark_lock);
298
299 list_for_each_entry_safe(mark, lmark, &free_list, free_g_list) {
300 fsnotify_destroy_mark(mark);
301 fsnotify_put_mark(mark);
302 }
303}
304
305/*
306 * Given a group, destroy all of the marks associated with that group.
307 */
308void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
309{
310 fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
311}
312
313void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
314{
315 assert_spin_locked(&old->lock);
316 new->i.inode = old->i.inode;
317 new->m.mnt = old->m.mnt;
318 new->group = old->group;
319 new->mask = old->mask;
320 new->free_mark = old->free_mark;
321}
322
323/*
324 * Nothing fancy, just initialize lists and locks and counters.
325 */
326void fsnotify_init_mark(struct fsnotify_mark *mark,
327 void (*free_mark)(struct fsnotify_mark *mark))
328{
329 memset(mark, 0, sizeof(*mark));
330 spin_lock_init(&mark->lock);
331 atomic_set(&mark->refcnt, 1);
332 mark->free_mark = free_mark;
333}
334
335static int fsnotify_mark_destroy(void *ignored)
336{
337 struct fsnotify_mark *mark, *next;
338 LIST_HEAD(private_destroy_list);
339
340 for (;;) {
341 spin_lock(&destroy_lock);
342 /* exchange the list head */
343 list_replace_init(&destroy_list, &private_destroy_list);
344 spin_unlock(&destroy_lock);
345
346 synchronize_srcu(&fsnotify_mark_srcu);
347
348 list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) {
349 list_del_init(&mark->destroy_list);
350 fsnotify_put_mark(mark);
351 }
352
353 wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
354 }
355
356 return 0;
357}
358
359static int __init fsnotify_mark_init(void)
360{
361 struct task_struct *thread;
362
363 thread = kthread_run(fsnotify_mark_destroy, NULL,
364 "fsnotify_mark");
365 if (IS_ERR(thread))
366 panic("unable to start fsnotify mark destruction thread.");
367
368 return 0;
369}
370device_initcall(fsnotify_mark_init);
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
4 */
5
6/*
7 * fsnotify inode mark locking/lifetime/and refcnting
8 *
9 * REFCNT:
10 * The group->recnt and mark->refcnt tell how many "things" in the kernel
11 * currently are referencing the objects. Both kind of objects typically will
12 * live inside the kernel with a refcnt of 2, one for its creation and one for
13 * the reference a group and a mark hold to each other.
14 * If you are holding the appropriate locks, you can take a reference and the
15 * object itself is guaranteed to survive until the reference is dropped.
16 *
17 * LOCKING:
18 * There are 3 locks involved with fsnotify inode marks and they MUST be taken
19 * in order as follows:
20 *
21 * group->mark_mutex
22 * mark->lock
23 * mark->connector->lock
24 *
25 * group->mark_mutex protects the marks_list anchored inside a given group and
26 * each mark is hooked via the g_list. It also protects the groups private
27 * data (i.e group limits).
28
29 * mark->lock protects the marks attributes like its masks and flags.
30 * Furthermore it protects the access to a reference of the group that the mark
31 * is assigned to as well as the access to a reference of the inode/vfsmount
32 * that is being watched by the mark.
33 *
34 * mark->connector->lock protects the list of marks anchored inside an
35 * inode / vfsmount and each mark is hooked via the i_list.
36 *
37 * A list of notification marks relating to inode / mnt is contained in
38 * fsnotify_mark_connector. That structure is alive as long as there are any
39 * marks in the list and is also protected by fsnotify_mark_srcu. A mark gets
40 * detached from fsnotify_mark_connector when last reference to the mark is
41 * dropped. Thus having mark reference is enough to protect mark->connector
42 * pointer and to make sure fsnotify_mark_connector cannot disappear. Also
43 * because we remove mark from g_list before dropping mark reference associated
44 * with that, any mark found through g_list is guaranteed to have
45 * mark->connector set until we drop group->mark_mutex.
46 *
47 * LIFETIME:
48 * Inode marks survive between when they are added to an inode and when their
49 * refcnt==0. Marks are also protected by fsnotify_mark_srcu.
50 *
51 * The inode mark can be cleared for a number of different reasons including:
52 * - The inode is unlinked for the last time. (fsnotify_inode_remove)
53 * - The inode is being evicted from cache. (fsnotify_inode_delete)
54 * - The fs the inode is on is unmounted. (fsnotify_inode_delete/fsnotify_unmount_inodes)
55 * - Something explicitly requests that it be removed. (fsnotify_destroy_mark)
56 * - The fsnotify_group associated with the mark is going away and all such marks
57 * need to be cleaned up. (fsnotify_clear_marks_by_group)
58 *
59 * This has the very interesting property of being able to run concurrently with
60 * any (or all) other directions.
61 */
62
63#include <linux/fs.h>
64#include <linux/init.h>
65#include <linux/kernel.h>
66#include <linux/kthread.h>
67#include <linux/module.h>
68#include <linux/mutex.h>
69#include <linux/slab.h>
70#include <linux/spinlock.h>
71#include <linux/srcu.h>
72#include <linux/ratelimit.h>
73
74#include <linux/atomic.h>
75
76#include <linux/fsnotify_backend.h>
77#include "fsnotify.h"
78
79#define FSNOTIFY_REAPER_DELAY (1) /* 1 jiffy */
80
81struct srcu_struct fsnotify_mark_srcu;
82struct kmem_cache *fsnotify_mark_connector_cachep;
83
84static DEFINE_SPINLOCK(destroy_lock);
85static LIST_HEAD(destroy_list);
86static struct fsnotify_mark_connector *connector_destroy_list;
87
88static void fsnotify_mark_destroy_workfn(struct work_struct *work);
89static DECLARE_DELAYED_WORK(reaper_work, fsnotify_mark_destroy_workfn);
90
91static void fsnotify_connector_destroy_workfn(struct work_struct *work);
92static DECLARE_WORK(connector_reaper_work, fsnotify_connector_destroy_workfn);
93
94void fsnotify_get_mark(struct fsnotify_mark *mark)
95{
96 WARN_ON_ONCE(!refcount_read(&mark->refcnt));
97 refcount_inc(&mark->refcnt);
98}
99
100static __u32 *fsnotify_conn_mask_p(struct fsnotify_mark_connector *conn)
101{
102 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
103 return &fsnotify_conn_inode(conn)->i_fsnotify_mask;
104 else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT)
105 return &fsnotify_conn_mount(conn)->mnt_fsnotify_mask;
106 else if (conn->type == FSNOTIFY_OBJ_TYPE_SB)
107 return &fsnotify_conn_sb(conn)->s_fsnotify_mask;
108 return NULL;
109}
110
111__u32 fsnotify_conn_mask(struct fsnotify_mark_connector *conn)
112{
113 if (WARN_ON(!fsnotify_valid_obj_type(conn->type)))
114 return 0;
115
116 return *fsnotify_conn_mask_p(conn);
117}
118
119static void __fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
120{
121 u32 new_mask = 0;
122 struct fsnotify_mark *mark;
123
124 assert_spin_locked(&conn->lock);
125 /* We can get detached connector here when inode is getting unlinked. */
126 if (!fsnotify_valid_obj_type(conn->type))
127 return;
128 hlist_for_each_entry(mark, &conn->list, obj_list) {
129 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)
130 new_mask |= mark->mask;
131 }
132 *fsnotify_conn_mask_p(conn) = new_mask;
133}
134
135/*
136 * Calculate mask of events for a list of marks. The caller must make sure
137 * connector and connector->obj cannot disappear under us. Callers achieve
138 * this by holding a mark->lock or mark->group->mark_mutex for a mark on this
139 * list.
140 */
141void fsnotify_recalc_mask(struct fsnotify_mark_connector *conn)
142{
143 if (!conn)
144 return;
145
146 spin_lock(&conn->lock);
147 __fsnotify_recalc_mask(conn);
148 spin_unlock(&conn->lock);
149 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
150 __fsnotify_update_child_dentry_flags(
151 fsnotify_conn_inode(conn));
152}
153
154/* Free all connectors queued for freeing once SRCU period ends */
155static void fsnotify_connector_destroy_workfn(struct work_struct *work)
156{
157 struct fsnotify_mark_connector *conn, *free;
158
159 spin_lock(&destroy_lock);
160 conn = connector_destroy_list;
161 connector_destroy_list = NULL;
162 spin_unlock(&destroy_lock);
163
164 synchronize_srcu(&fsnotify_mark_srcu);
165 while (conn) {
166 free = conn;
167 conn = conn->destroy_next;
168 kmem_cache_free(fsnotify_mark_connector_cachep, free);
169 }
170}
171
172static void *fsnotify_detach_connector_from_object(
173 struct fsnotify_mark_connector *conn,
174 unsigned int *type)
175{
176 struct inode *inode = NULL;
177
178 *type = conn->type;
179 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED)
180 return NULL;
181
182 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE) {
183 inode = fsnotify_conn_inode(conn);
184 inode->i_fsnotify_mask = 0;
185 atomic_long_inc(&inode->i_sb->s_fsnotify_inode_refs);
186 } else if (conn->type == FSNOTIFY_OBJ_TYPE_VFSMOUNT) {
187 fsnotify_conn_mount(conn)->mnt_fsnotify_mask = 0;
188 } else if (conn->type == FSNOTIFY_OBJ_TYPE_SB) {
189 fsnotify_conn_sb(conn)->s_fsnotify_mask = 0;
190 }
191
192 rcu_assign_pointer(*(conn->obj), NULL);
193 conn->obj = NULL;
194 conn->type = FSNOTIFY_OBJ_TYPE_DETACHED;
195
196 return inode;
197}
198
199static void fsnotify_final_mark_destroy(struct fsnotify_mark *mark)
200{
201 struct fsnotify_group *group = mark->group;
202
203 if (WARN_ON_ONCE(!group))
204 return;
205 group->ops->free_mark(mark);
206 fsnotify_put_group(group);
207}
208
209/* Drop object reference originally held by a connector */
210static void fsnotify_drop_object(unsigned int type, void *objp)
211{
212 struct inode *inode;
213 struct super_block *sb;
214
215 if (!objp)
216 return;
217 /* Currently only inode references are passed to be dropped */
218 if (WARN_ON_ONCE(type != FSNOTIFY_OBJ_TYPE_INODE))
219 return;
220 inode = objp;
221 sb = inode->i_sb;
222 iput(inode);
223 if (atomic_long_dec_and_test(&sb->s_fsnotify_inode_refs))
224 wake_up_var(&sb->s_fsnotify_inode_refs);
225}
226
227void fsnotify_put_mark(struct fsnotify_mark *mark)
228{
229 struct fsnotify_mark_connector *conn = READ_ONCE(mark->connector);
230 void *objp = NULL;
231 unsigned int type = FSNOTIFY_OBJ_TYPE_DETACHED;
232 bool free_conn = false;
233
234 /* Catch marks that were actually never attached to object */
235 if (!conn) {
236 if (refcount_dec_and_test(&mark->refcnt))
237 fsnotify_final_mark_destroy(mark);
238 return;
239 }
240
241 /*
242 * We have to be careful so that traversals of obj_list under lock can
243 * safely grab mark reference.
244 */
245 if (!refcount_dec_and_lock(&mark->refcnt, &conn->lock))
246 return;
247
248 hlist_del_init_rcu(&mark->obj_list);
249 if (hlist_empty(&conn->list)) {
250 objp = fsnotify_detach_connector_from_object(conn, &type);
251 free_conn = true;
252 } else {
253 __fsnotify_recalc_mask(conn);
254 }
255 WRITE_ONCE(mark->connector, NULL);
256 spin_unlock(&conn->lock);
257
258 fsnotify_drop_object(type, objp);
259
260 if (free_conn) {
261 spin_lock(&destroy_lock);
262 conn->destroy_next = connector_destroy_list;
263 connector_destroy_list = conn;
264 spin_unlock(&destroy_lock);
265 queue_work(system_unbound_wq, &connector_reaper_work);
266 }
267 /*
268 * Note that we didn't update flags telling whether inode cares about
269 * what's happening with children. We update these flags from
270 * __fsnotify_parent() lazily when next event happens on one of our
271 * children.
272 */
273 spin_lock(&destroy_lock);
274 list_add(&mark->g_list, &destroy_list);
275 spin_unlock(&destroy_lock);
276 queue_delayed_work(system_unbound_wq, &reaper_work,
277 FSNOTIFY_REAPER_DELAY);
278}
279EXPORT_SYMBOL_GPL(fsnotify_put_mark);
280
281/*
282 * Get mark reference when we found the mark via lockless traversal of object
283 * list. Mark can be already removed from the list by now and on its way to be
284 * destroyed once SRCU period ends.
285 *
286 * Also pin the group so it doesn't disappear under us.
287 */
288static bool fsnotify_get_mark_safe(struct fsnotify_mark *mark)
289{
290 if (!mark)
291 return true;
292
293 if (refcount_inc_not_zero(&mark->refcnt)) {
294 spin_lock(&mark->lock);
295 if (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) {
296 /* mark is attached, group is still alive then */
297 atomic_inc(&mark->group->user_waits);
298 spin_unlock(&mark->lock);
299 return true;
300 }
301 spin_unlock(&mark->lock);
302 fsnotify_put_mark(mark);
303 }
304 return false;
305}
306
307/*
308 * Puts marks and wakes up group destruction if necessary.
309 *
310 * Pairs with fsnotify_get_mark_safe()
311 */
312static void fsnotify_put_mark_wake(struct fsnotify_mark *mark)
313{
314 if (mark) {
315 struct fsnotify_group *group = mark->group;
316
317 fsnotify_put_mark(mark);
318 /*
319 * We abuse notification_waitq on group shutdown for waiting for
320 * all marks pinned when waiting for userspace.
321 */
322 if (atomic_dec_and_test(&group->user_waits) && group->shutdown)
323 wake_up(&group->notification_waitq);
324 }
325}
326
327bool fsnotify_prepare_user_wait(struct fsnotify_iter_info *iter_info)
328{
329 int type;
330
331 fsnotify_foreach_obj_type(type) {
332 /* This can fail if mark is being removed */
333 if (!fsnotify_get_mark_safe(iter_info->marks[type]))
334 goto fail;
335 }
336
337 /*
338 * Now that both marks are pinned by refcount in the inode / vfsmount
339 * lists, we can drop SRCU lock, and safely resume the list iteration
340 * once userspace returns.
341 */
342 srcu_read_unlock(&fsnotify_mark_srcu, iter_info->srcu_idx);
343
344 return true;
345
346fail:
347 for (type--; type >= 0; type--)
348 fsnotify_put_mark_wake(iter_info->marks[type]);
349 return false;
350}
351
352void fsnotify_finish_user_wait(struct fsnotify_iter_info *iter_info)
353{
354 int type;
355
356 iter_info->srcu_idx = srcu_read_lock(&fsnotify_mark_srcu);
357 fsnotify_foreach_obj_type(type)
358 fsnotify_put_mark_wake(iter_info->marks[type]);
359}
360
361/*
362 * Mark mark as detached, remove it from group list. Mark still stays in object
363 * list until its last reference is dropped. Note that we rely on mark being
364 * removed from group list before corresponding reference to it is dropped. In
365 * particular we rely on mark->connector being valid while we hold
366 * group->mark_mutex if we found the mark through g_list.
367 *
368 * Must be called with group->mark_mutex held. The caller must either hold
369 * reference to the mark or be protected by fsnotify_mark_srcu.
370 */
371void fsnotify_detach_mark(struct fsnotify_mark *mark)
372{
373 struct fsnotify_group *group = mark->group;
374
375 WARN_ON_ONCE(!mutex_is_locked(&group->mark_mutex));
376 WARN_ON_ONCE(!srcu_read_lock_held(&fsnotify_mark_srcu) &&
377 refcount_read(&mark->refcnt) < 1 +
378 !!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED));
379
380 spin_lock(&mark->lock);
381 /* something else already called this function on this mark */
382 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
383 spin_unlock(&mark->lock);
384 return;
385 }
386 mark->flags &= ~FSNOTIFY_MARK_FLAG_ATTACHED;
387 list_del_init(&mark->g_list);
388 spin_unlock(&mark->lock);
389
390 atomic_dec(&group->num_marks);
391
392 /* Drop mark reference acquired in fsnotify_add_mark_locked() */
393 fsnotify_put_mark(mark);
394}
395
396/*
397 * Free fsnotify mark. The mark is actually only marked as being freed. The
398 * freeing is actually happening only once last reference to the mark is
399 * dropped from a workqueue which first waits for srcu period end.
400 *
401 * Caller must have a reference to the mark or be protected by
402 * fsnotify_mark_srcu.
403 */
404void fsnotify_free_mark(struct fsnotify_mark *mark)
405{
406 struct fsnotify_group *group = mark->group;
407
408 spin_lock(&mark->lock);
409 /* something else already called this function on this mark */
410 if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
411 spin_unlock(&mark->lock);
412 return;
413 }
414 mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
415 spin_unlock(&mark->lock);
416
417 /*
418 * Some groups like to know that marks are being freed. This is a
419 * callback to the group function to let it know that this mark
420 * is being freed.
421 */
422 if (group->ops->freeing_mark)
423 group->ops->freeing_mark(mark, group);
424}
425
426void fsnotify_destroy_mark(struct fsnotify_mark *mark,
427 struct fsnotify_group *group)
428{
429 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
430 fsnotify_detach_mark(mark);
431 mutex_unlock(&group->mark_mutex);
432 fsnotify_free_mark(mark);
433}
434EXPORT_SYMBOL_GPL(fsnotify_destroy_mark);
435
436/*
437 * Sorting function for lists of fsnotify marks.
438 *
439 * Fanotify supports different notification classes (reflected as priority of
440 * notification group). Events shall be passed to notification groups in
441 * decreasing priority order. To achieve this marks in notification lists for
442 * inodes and vfsmounts are sorted so that priorities of corresponding groups
443 * are descending.
444 *
445 * Furthermore correct handling of the ignore mask requires processing inode
446 * and vfsmount marks of each group together. Using the group address as
447 * further sort criterion provides a unique sorting order and thus we can
448 * merge inode and vfsmount lists of marks in linear time and find groups
449 * present in both lists.
450 *
451 * A return value of 1 signifies that b has priority over a.
452 * A return value of 0 signifies that the two marks have to be handled together.
453 * A return value of -1 signifies that a has priority over b.
454 */
455int fsnotify_compare_groups(struct fsnotify_group *a, struct fsnotify_group *b)
456{
457 if (a == b)
458 return 0;
459 if (!a)
460 return 1;
461 if (!b)
462 return -1;
463 if (a->priority < b->priority)
464 return 1;
465 if (a->priority > b->priority)
466 return -1;
467 if (a < b)
468 return 1;
469 return -1;
470}
471
472static int fsnotify_attach_connector_to_object(fsnotify_connp_t *connp,
473 unsigned int type,
474 __kernel_fsid_t *fsid)
475{
476 struct inode *inode = NULL;
477 struct fsnotify_mark_connector *conn;
478
479 conn = kmem_cache_alloc(fsnotify_mark_connector_cachep, GFP_KERNEL);
480 if (!conn)
481 return -ENOMEM;
482 spin_lock_init(&conn->lock);
483 INIT_HLIST_HEAD(&conn->list);
484 conn->type = type;
485 conn->obj = connp;
486 /* Cache fsid of filesystem containing the object */
487 if (fsid) {
488 conn->fsid = *fsid;
489 conn->flags = FSNOTIFY_CONN_FLAG_HAS_FSID;
490 } else {
491 conn->fsid.val[0] = conn->fsid.val[1] = 0;
492 conn->flags = 0;
493 }
494 if (conn->type == FSNOTIFY_OBJ_TYPE_INODE)
495 inode = igrab(fsnotify_conn_inode(conn));
496 /*
497 * cmpxchg() provides the barrier so that readers of *connp can see
498 * only initialized structure
499 */
500 if (cmpxchg(connp, NULL, conn)) {
501 /* Someone else created list structure for us */
502 if (inode)
503 iput(inode);
504 kmem_cache_free(fsnotify_mark_connector_cachep, conn);
505 }
506
507 return 0;
508}
509
510/*
511 * Get mark connector, make sure it is alive and return with its lock held.
512 * This is for users that get connector pointer from inode or mount. Users that
513 * hold reference to a mark on the list may directly lock connector->lock as
514 * they are sure list cannot go away under them.
515 */
516static struct fsnotify_mark_connector *fsnotify_grab_connector(
517 fsnotify_connp_t *connp)
518{
519 struct fsnotify_mark_connector *conn;
520 int idx;
521
522 idx = srcu_read_lock(&fsnotify_mark_srcu);
523 conn = srcu_dereference(*connp, &fsnotify_mark_srcu);
524 if (!conn)
525 goto out;
526 spin_lock(&conn->lock);
527 if (conn->type == FSNOTIFY_OBJ_TYPE_DETACHED) {
528 spin_unlock(&conn->lock);
529 srcu_read_unlock(&fsnotify_mark_srcu, idx);
530 return NULL;
531 }
532out:
533 srcu_read_unlock(&fsnotify_mark_srcu, idx);
534 return conn;
535}
536
537/*
538 * Add mark into proper place in given list of marks. These marks may be used
539 * for the fsnotify backend to determine which event types should be delivered
540 * to which group and for which inodes. These marks are ordered according to
541 * priority, highest number first, and then by the group's location in memory.
542 */
543static int fsnotify_add_mark_list(struct fsnotify_mark *mark,
544 fsnotify_connp_t *connp, unsigned int type,
545 int allow_dups, __kernel_fsid_t *fsid)
546{
547 struct fsnotify_mark *lmark, *last = NULL;
548 struct fsnotify_mark_connector *conn;
549 int cmp;
550 int err = 0;
551
552 if (WARN_ON(!fsnotify_valid_obj_type(type)))
553 return -EINVAL;
554
555 /* Backend is expected to check for zero fsid (e.g. tmpfs) */
556 if (fsid && WARN_ON_ONCE(!fsid->val[0] && !fsid->val[1]))
557 return -ENODEV;
558
559restart:
560 spin_lock(&mark->lock);
561 conn = fsnotify_grab_connector(connp);
562 if (!conn) {
563 spin_unlock(&mark->lock);
564 err = fsnotify_attach_connector_to_object(connp, type, fsid);
565 if (err)
566 return err;
567 goto restart;
568 } else if (fsid && !(conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID)) {
569 conn->fsid = *fsid;
570 /* Pairs with smp_rmb() in fanotify_get_fsid() */
571 smp_wmb();
572 conn->flags |= FSNOTIFY_CONN_FLAG_HAS_FSID;
573 } else if (fsid && (conn->flags & FSNOTIFY_CONN_FLAG_HAS_FSID) &&
574 (fsid->val[0] != conn->fsid.val[0] ||
575 fsid->val[1] != conn->fsid.val[1])) {
576 /*
577 * Backend is expected to check for non uniform fsid
578 * (e.g. btrfs), but maybe we missed something?
579 * Only allow setting conn->fsid once to non zero fsid.
580 * inotify and non-fid fanotify groups do not set nor test
581 * conn->fsid.
582 */
583 pr_warn_ratelimited("%s: fsid mismatch on object of type %u: "
584 "%x.%x != %x.%x\n", __func__, conn->type,
585 fsid->val[0], fsid->val[1],
586 conn->fsid.val[0], conn->fsid.val[1]);
587 err = -EXDEV;
588 goto out_err;
589 }
590
591 /* is mark the first mark? */
592 if (hlist_empty(&conn->list)) {
593 hlist_add_head_rcu(&mark->obj_list, &conn->list);
594 goto added;
595 }
596
597 /* should mark be in the middle of the current list? */
598 hlist_for_each_entry(lmark, &conn->list, obj_list) {
599 last = lmark;
600
601 if ((lmark->group == mark->group) &&
602 (lmark->flags & FSNOTIFY_MARK_FLAG_ATTACHED) &&
603 !allow_dups) {
604 err = -EEXIST;
605 goto out_err;
606 }
607
608 cmp = fsnotify_compare_groups(lmark->group, mark->group);
609 if (cmp >= 0) {
610 hlist_add_before_rcu(&mark->obj_list, &lmark->obj_list);
611 goto added;
612 }
613 }
614
615 BUG_ON(last == NULL);
616 /* mark should be the last entry. last is the current last entry */
617 hlist_add_behind_rcu(&mark->obj_list, &last->obj_list);
618added:
619 /*
620 * Since connector is attached to object using cmpxchg() we are
621 * guaranteed that connector initialization is fully visible by anyone
622 * seeing mark->connector set.
623 */
624 WRITE_ONCE(mark->connector, conn);
625out_err:
626 spin_unlock(&conn->lock);
627 spin_unlock(&mark->lock);
628 return err;
629}
630
631/*
632 * Attach an initialized mark to a given group and fs object.
633 * These marks may be used for the fsnotify backend to determine which
634 * event types should be delivered to which group.
635 */
636int fsnotify_add_mark_locked(struct fsnotify_mark *mark,
637 fsnotify_connp_t *connp, unsigned int type,
638 int allow_dups, __kernel_fsid_t *fsid)
639{
640 struct fsnotify_group *group = mark->group;
641 int ret = 0;
642
643 BUG_ON(!mutex_is_locked(&group->mark_mutex));
644
645 /*
646 * LOCKING ORDER!!!!
647 * group->mark_mutex
648 * mark->lock
649 * mark->connector->lock
650 */
651 spin_lock(&mark->lock);
652 mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE | FSNOTIFY_MARK_FLAG_ATTACHED;
653
654 list_add(&mark->g_list, &group->marks_list);
655 atomic_inc(&group->num_marks);
656 fsnotify_get_mark(mark); /* for g_list */
657 spin_unlock(&mark->lock);
658
659 ret = fsnotify_add_mark_list(mark, connp, type, allow_dups, fsid);
660 if (ret)
661 goto err;
662
663 if (mark->mask)
664 fsnotify_recalc_mask(mark->connector);
665
666 return ret;
667err:
668 spin_lock(&mark->lock);
669 mark->flags &= ~(FSNOTIFY_MARK_FLAG_ALIVE |
670 FSNOTIFY_MARK_FLAG_ATTACHED);
671 list_del_init(&mark->g_list);
672 spin_unlock(&mark->lock);
673 atomic_dec(&group->num_marks);
674
675 fsnotify_put_mark(mark);
676 return ret;
677}
678
679int fsnotify_add_mark(struct fsnotify_mark *mark, fsnotify_connp_t *connp,
680 unsigned int type, int allow_dups, __kernel_fsid_t *fsid)
681{
682 int ret;
683 struct fsnotify_group *group = mark->group;
684
685 mutex_lock(&group->mark_mutex);
686 ret = fsnotify_add_mark_locked(mark, connp, type, allow_dups, fsid);
687 mutex_unlock(&group->mark_mutex);
688 return ret;
689}
690EXPORT_SYMBOL_GPL(fsnotify_add_mark);
691
692/*
693 * Given a list of marks, find the mark associated with given group. If found
694 * take a reference to that mark and return it, else return NULL.
695 */
696struct fsnotify_mark *fsnotify_find_mark(fsnotify_connp_t *connp,
697 struct fsnotify_group *group)
698{
699 struct fsnotify_mark_connector *conn;
700 struct fsnotify_mark *mark;
701
702 conn = fsnotify_grab_connector(connp);
703 if (!conn)
704 return NULL;
705
706 hlist_for_each_entry(mark, &conn->list, obj_list) {
707 if (mark->group == group &&
708 (mark->flags & FSNOTIFY_MARK_FLAG_ATTACHED)) {
709 fsnotify_get_mark(mark);
710 spin_unlock(&conn->lock);
711 return mark;
712 }
713 }
714 spin_unlock(&conn->lock);
715 return NULL;
716}
717EXPORT_SYMBOL_GPL(fsnotify_find_mark);
718
719/* Clear any marks in a group with given type mask */
720void fsnotify_clear_marks_by_group(struct fsnotify_group *group,
721 unsigned int type_mask)
722{
723 struct fsnotify_mark *lmark, *mark;
724 LIST_HEAD(to_free);
725 struct list_head *head = &to_free;
726
727 /* Skip selection step if we want to clear all marks. */
728 if (type_mask == FSNOTIFY_OBJ_ALL_TYPES_MASK) {
729 head = &group->marks_list;
730 goto clear;
731 }
732 /*
733 * We have to be really careful here. Anytime we drop mark_mutex, e.g.
734 * fsnotify_clear_marks_by_inode() can come and free marks. Even in our
735 * to_free list so we have to use mark_mutex even when accessing that
736 * list. And freeing mark requires us to drop mark_mutex. So we can
737 * reliably free only the first mark in the list. That's why we first
738 * move marks to free to to_free list in one go and then free marks in
739 * to_free list one by one.
740 */
741 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
742 list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
743 if ((1U << mark->connector->type) & type_mask)
744 list_move(&mark->g_list, &to_free);
745 }
746 mutex_unlock(&group->mark_mutex);
747
748clear:
749 while (1) {
750 mutex_lock_nested(&group->mark_mutex, SINGLE_DEPTH_NESTING);
751 if (list_empty(head)) {
752 mutex_unlock(&group->mark_mutex);
753 break;
754 }
755 mark = list_first_entry(head, struct fsnotify_mark, g_list);
756 fsnotify_get_mark(mark);
757 fsnotify_detach_mark(mark);
758 mutex_unlock(&group->mark_mutex);
759 fsnotify_free_mark(mark);
760 fsnotify_put_mark(mark);
761 }
762}
763
764/* Destroy all marks attached to an object via connector */
765void fsnotify_destroy_marks(fsnotify_connp_t *connp)
766{
767 struct fsnotify_mark_connector *conn;
768 struct fsnotify_mark *mark, *old_mark = NULL;
769 void *objp;
770 unsigned int type;
771
772 conn = fsnotify_grab_connector(connp);
773 if (!conn)
774 return;
775 /*
776 * We have to be careful since we can race with e.g.
777 * fsnotify_clear_marks_by_group() and once we drop the conn->lock, the
778 * list can get modified. However we are holding mark reference and
779 * thus our mark cannot be removed from obj_list so we can continue
780 * iteration after regaining conn->lock.
781 */
782 hlist_for_each_entry(mark, &conn->list, obj_list) {
783 fsnotify_get_mark(mark);
784 spin_unlock(&conn->lock);
785 if (old_mark)
786 fsnotify_put_mark(old_mark);
787 old_mark = mark;
788 fsnotify_destroy_mark(mark, mark->group);
789 spin_lock(&conn->lock);
790 }
791 /*
792 * Detach list from object now so that we don't pin inode until all
793 * mark references get dropped. It would lead to strange results such
794 * as delaying inode deletion or blocking unmount.
795 */
796 objp = fsnotify_detach_connector_from_object(conn, &type);
797 spin_unlock(&conn->lock);
798 if (old_mark)
799 fsnotify_put_mark(old_mark);
800 fsnotify_drop_object(type, objp);
801}
802
803/*
804 * Nothing fancy, just initialize lists and locks and counters.
805 */
806void fsnotify_init_mark(struct fsnotify_mark *mark,
807 struct fsnotify_group *group)
808{
809 memset(mark, 0, sizeof(*mark));
810 spin_lock_init(&mark->lock);
811 refcount_set(&mark->refcnt, 1);
812 fsnotify_get_group(group);
813 mark->group = group;
814 WRITE_ONCE(mark->connector, NULL);
815}
816EXPORT_SYMBOL_GPL(fsnotify_init_mark);
817
818/*
819 * Destroy all marks in destroy_list, waits for SRCU period to finish before
820 * actually freeing marks.
821 */
822static void fsnotify_mark_destroy_workfn(struct work_struct *work)
823{
824 struct fsnotify_mark *mark, *next;
825 struct list_head private_destroy_list;
826
827 spin_lock(&destroy_lock);
828 /* exchange the list head */
829 list_replace_init(&destroy_list, &private_destroy_list);
830 spin_unlock(&destroy_lock);
831
832 synchronize_srcu(&fsnotify_mark_srcu);
833
834 list_for_each_entry_safe(mark, next, &private_destroy_list, g_list) {
835 list_del_init(&mark->g_list);
836 fsnotify_final_mark_destroy(mark);
837 }
838}
839
840/* Wait for all marks queued for destruction to be actually destroyed */
841void fsnotify_wait_marks_destroyed(void)
842{
843 flush_delayed_work(&reaper_work);
844}
845EXPORT_SYMBOL_GPL(fsnotify_wait_marks_destroyed);