blob: 282150f74cfaaf367eaaa6a5dc4fcfd845760f41 [file] [log] [blame]
Eric Paris3be25f42009-05-21 17:01:26 -04001/*
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 guarenteed 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 * entry->lock
34 * group->mark_lock
35 * inode->i_lock
36 *
37 * entry->lock protects 2 things, entry->group and entry->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 mark_entries list anchored inside a given group
42 * and each entry 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_mark_entries list anchored inside a
47 * given inode and each entry 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_by_entry)
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_mark_entries 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/module.h>
89#include <linux/mutex.h>
90#include <linux/slab.h>
91#include <linux/spinlock.h>
92
93#include <asm/atomic.h>
94
95#include <linux/fsnotify_backend.h>
96#include "fsnotify.h"
97
98void fsnotify_get_mark(struct fsnotify_mark_entry *entry)
99{
100 atomic_inc(&entry->refcnt);
101}
102
103void fsnotify_put_mark(struct fsnotify_mark_entry *entry)
104{
105 if (atomic_dec_and_test(&entry->refcnt))
106 entry->free_mark(entry);
107}
108
109/*
110 * Recalculate the mask of events relevant to a given inode locked.
111 */
112static void fsnotify_recalc_inode_mask_locked(struct inode *inode)
113{
114 struct fsnotify_mark_entry *entry;
115 struct hlist_node *pos;
116 __u32 new_mask = 0;
117
118 assert_spin_locked(&inode->i_lock);
119
120 hlist_for_each_entry(entry, pos, &inode->i_fsnotify_mark_entries, i_list)
121 new_mask |= entry->mask;
122 inode->i_fsnotify_mask = new_mask;
123}
124
125/*
126 * Recalculate the inode->i_fsnotify_mask, or the mask of all FS_* event types
127 * any notifier is interested in hearing for this inode.
128 */
129void fsnotify_recalc_inode_mask(struct inode *inode)
130{
131 spin_lock(&inode->i_lock);
132 fsnotify_recalc_inode_mask_locked(inode);
133 spin_unlock(&inode->i_lock);
Eric Parisc28f7e52009-05-21 17:01:29 -0400134
135 __fsnotify_update_child_dentry_flags(inode);
Eric Paris3be25f42009-05-21 17:01:26 -0400136}
137
138/*
139 * Any time a mark is getting freed we end up here.
140 * The caller had better be holding a reference to this mark so we don't actually
141 * do the final put under the entry->lock
142 */
143void fsnotify_destroy_mark_by_entry(struct fsnotify_mark_entry *entry)
144{
145 struct fsnotify_group *group;
146 struct inode *inode;
147
148 spin_lock(&entry->lock);
149
150 group = entry->group;
151 inode = entry->inode;
152
153 BUG_ON(group && !inode);
154 BUG_ON(!group && inode);
155
156 /* if !group something else already marked this to die */
157 if (!group) {
158 spin_unlock(&entry->lock);
159 return;
160 }
161
162 /* 1 from caller and 1 for being on i_list/g_list */
163 BUG_ON(atomic_read(&entry->refcnt) < 2);
164
165 spin_lock(&group->mark_lock);
166 spin_lock(&inode->i_lock);
167
168 hlist_del_init(&entry->i_list);
169 entry->inode = NULL;
170
171 list_del_init(&entry->g_list);
172 entry->group = NULL;
173
174 fsnotify_put_mark(entry); /* for i_list and g_list */
175
176 /*
177 * this mark is now off the inode->i_fsnotify_mark_entries list and we
178 * hold the inode->i_lock, so this is the perfect time to update the
179 * inode->i_fsnotify_mask
180 */
181 fsnotify_recalc_inode_mask_locked(inode);
182
183 spin_unlock(&inode->i_lock);
184 spin_unlock(&group->mark_lock);
185 spin_unlock(&entry->lock);
186
187 /*
188 * Some groups like to know that marks are being freed. This is a
189 * callback to the group function to let it know that this entry
190 * is being freed.
191 */
192 group->ops->freeing_mark(entry, group);
193
194 /*
Eric Parisc28f7e52009-05-21 17:01:29 -0400195 * __fsnotify_update_child_dentry_flags(inode);
196 *
197 * I really want to call that, but we can't, we have no idea if the inode
198 * still exists the second we drop the entry->lock.
199 *
200 * The next time an event arrive to this inode from one of it's children
201 * __fsnotify_parent will see that the inode doesn't care about it's
202 * children and will update all of these flags then. So really this
203 * is just a lazy update (and could be a perf win...)
204 */
205
206
Eric Paris1ef5f132009-05-21 17:01:54 -0400207 iput(inode);
208
Eric Parisc28f7e52009-05-21 17:01:29 -0400209 /*
Eric Paris3be25f42009-05-21 17:01:26 -0400210 * it's possible that this group tried to destroy itself, but this
211 * this mark was simultaneously being freed by inode. If that's the
212 * case, we finish freeing the group here.
213 */
214 if (unlikely(atomic_dec_and_test(&group->num_marks)))
215 fsnotify_final_destroy_group(group);
216}
217
218/*
219 * Given a group, destroy all of the marks associated with that group.
220 */
221void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
222{
223 struct fsnotify_mark_entry *lentry, *entry;
224 LIST_HEAD(free_list);
225
226 spin_lock(&group->mark_lock);
227 list_for_each_entry_safe(entry, lentry, &group->mark_entries, g_list) {
228 list_add(&entry->free_g_list, &free_list);
229 list_del_init(&entry->g_list);
230 fsnotify_get_mark(entry);
231 }
232 spin_unlock(&group->mark_lock);
233
234 list_for_each_entry_safe(entry, lentry, &free_list, free_g_list) {
235 fsnotify_destroy_mark_by_entry(entry);
236 fsnotify_put_mark(entry);
237 }
238}
239
240/*
241 * Given an inode, destroy all of the marks associated with that inode.
242 */
243void fsnotify_clear_marks_by_inode(struct inode *inode)
244{
245 struct fsnotify_mark_entry *entry, *lentry;
246 struct hlist_node *pos, *n;
247 LIST_HEAD(free_list);
248
249 spin_lock(&inode->i_lock);
250 hlist_for_each_entry_safe(entry, pos, n, &inode->i_fsnotify_mark_entries, i_list) {
251 list_add(&entry->free_i_list, &free_list);
252 hlist_del_init(&entry->i_list);
253 fsnotify_get_mark(entry);
254 }
255 spin_unlock(&inode->i_lock);
256
257 list_for_each_entry_safe(entry, lentry, &free_list, free_i_list) {
258 fsnotify_destroy_mark_by_entry(entry);
259 fsnotify_put_mark(entry);
260 }
261}
262
263/*
264 * given a group and inode, find the mark associated with that combination.
265 * if found take a reference to that mark and return it, else return NULL
266 */
267struct fsnotify_mark_entry *fsnotify_find_mark_entry(struct fsnotify_group *group,
268 struct inode *inode)
269{
270 struct fsnotify_mark_entry *entry;
271 struct hlist_node *pos;
272
273 assert_spin_locked(&inode->i_lock);
274
275 hlist_for_each_entry(entry, pos, &inode->i_fsnotify_mark_entries, i_list) {
276 if (entry->group == group) {
277 fsnotify_get_mark(entry);
278 return entry;
279 }
280 }
281 return NULL;
282}
283
284/*
285 * Nothing fancy, just initialize lists and locks and counters.
286 */
287void fsnotify_init_mark(struct fsnotify_mark_entry *entry,
288 void (*free_mark)(struct fsnotify_mark_entry *entry))
289
290{
291 spin_lock_init(&entry->lock);
292 atomic_set(&entry->refcnt, 1);
293 INIT_HLIST_NODE(&entry->i_list);
294 entry->group = NULL;
295 entry->mask = 0;
296 entry->inode = NULL;
297 entry->free_mark = free_mark;
298}
299
300/*
301 * Attach an initialized mark entry to a given group and inode.
302 * These marks may be used for the fsnotify backend to determine which
303 * event types should be delivered to which group and for which inodes.
304 */
305int fsnotify_add_mark(struct fsnotify_mark_entry *entry,
306 struct fsnotify_group *group, struct inode *inode)
307{
308 struct fsnotify_mark_entry *lentry;
309 int ret = 0;
310
Eric Paris1ef5f132009-05-21 17:01:54 -0400311 inode = igrab(inode);
312 if (unlikely(!inode))
313 return -EINVAL;
314
Eric Paris3be25f42009-05-21 17:01:26 -0400315 /*
316 * LOCKING ORDER!!!!
317 * entry->lock
318 * group->mark_lock
319 * inode->i_lock
320 */
321 spin_lock(&entry->lock);
322 spin_lock(&group->mark_lock);
323 spin_lock(&inode->i_lock);
324
325 entry->group = group;
326 entry->inode = inode;
327
328 lentry = fsnotify_find_mark_entry(group, inode);
329 if (!lentry) {
330 hlist_add_head(&entry->i_list, &inode->i_fsnotify_mark_entries);
331 list_add(&entry->g_list, &group->mark_entries);
332
333 fsnotify_get_mark(entry); /* for i_list and g_list */
334
335 atomic_inc(&group->num_marks);
336
337 fsnotify_recalc_inode_mask_locked(inode);
338 }
339
340 spin_unlock(&inode->i_lock);
341 spin_unlock(&group->mark_lock);
342 spin_unlock(&entry->lock);
343
344 if (lentry) {
345 ret = -EEXIST;
Eric Paris1ef5f132009-05-21 17:01:54 -0400346 iput(inode);
Eric Paris3be25f42009-05-21 17:01:26 -0400347 fsnotify_put_mark(lentry);
Eric Parisc28f7e52009-05-21 17:01:29 -0400348 } else {
349 __fsnotify_update_child_dentry_flags(inode);
Eric Paris3be25f42009-05-21 17:01:26 -0400350 }
351
352 return ret;
353}