]>
Commit | Line | Data |
---|---|---|
3be25f49 EP |
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 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 | ||
98 | void fsnotify_get_mark(struct fsnotify_mark_entry *entry) | |
99 | { | |
100 | atomic_inc(&entry->refcnt); | |
101 | } | |
102 | ||
103 | void 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 | */ | |
112 | static 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 | */ | |
129 | void 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); | |
c28f7e56 EP |
134 | |
135 | __fsnotify_update_child_dentry_flags(inode); | |
3be25f49 EP |
136 | } |
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 | */ | |
143 | void 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 | ||
c28f7e56 EP |
194 | /* |
195 | * __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 | ||
3be25f49 EP |
207 | /* |
208 | * it's possible that this group tried to destroy itself, but this | |
209 | * this mark was simultaneously being freed by inode. If that's the | |
210 | * case, we finish freeing the group here. | |
211 | */ | |
212 | if (unlikely(atomic_dec_and_test(&group->num_marks))) | |
213 | fsnotify_final_destroy_group(group); | |
214 | } | |
215 | ||
216 | /* | |
217 | * Given a group, destroy all of the marks associated with that group. | |
218 | */ | |
219 | void fsnotify_clear_marks_by_group(struct fsnotify_group *group) | |
220 | { | |
221 | struct fsnotify_mark_entry *lentry, *entry; | |
222 | LIST_HEAD(free_list); | |
223 | ||
224 | spin_lock(&group->mark_lock); | |
225 | list_for_each_entry_safe(entry, lentry, &group->mark_entries, g_list) { | |
226 | list_add(&entry->free_g_list, &free_list); | |
227 | list_del_init(&entry->g_list); | |
228 | fsnotify_get_mark(entry); | |
229 | } | |
230 | spin_unlock(&group->mark_lock); | |
231 | ||
232 | list_for_each_entry_safe(entry, lentry, &free_list, free_g_list) { | |
233 | fsnotify_destroy_mark_by_entry(entry); | |
234 | fsnotify_put_mark(entry); | |
235 | } | |
236 | } | |
237 | ||
238 | /* | |
239 | * Given an inode, destroy all of the marks associated with that inode. | |
240 | */ | |
241 | void fsnotify_clear_marks_by_inode(struct inode *inode) | |
242 | { | |
243 | struct fsnotify_mark_entry *entry, *lentry; | |
244 | struct hlist_node *pos, *n; | |
245 | LIST_HEAD(free_list); | |
246 | ||
247 | spin_lock(&inode->i_lock); | |
248 | hlist_for_each_entry_safe(entry, pos, n, &inode->i_fsnotify_mark_entries, i_list) { | |
249 | list_add(&entry->free_i_list, &free_list); | |
250 | hlist_del_init(&entry->i_list); | |
251 | fsnotify_get_mark(entry); | |
252 | } | |
253 | spin_unlock(&inode->i_lock); | |
254 | ||
255 | list_for_each_entry_safe(entry, lentry, &free_list, free_i_list) { | |
256 | fsnotify_destroy_mark_by_entry(entry); | |
257 | fsnotify_put_mark(entry); | |
258 | } | |
259 | } | |
260 | ||
261 | /* | |
262 | * given a group and inode, find the mark associated with that combination. | |
263 | * if found take a reference to that mark and return it, else return NULL | |
264 | */ | |
265 | struct fsnotify_mark_entry *fsnotify_find_mark_entry(struct fsnotify_group *group, | |
266 | struct inode *inode) | |
267 | { | |
268 | struct fsnotify_mark_entry *entry; | |
269 | struct hlist_node *pos; | |
270 | ||
271 | assert_spin_locked(&inode->i_lock); | |
272 | ||
273 | hlist_for_each_entry(entry, pos, &inode->i_fsnotify_mark_entries, i_list) { | |
274 | if (entry->group == group) { | |
275 | fsnotify_get_mark(entry); | |
276 | return entry; | |
277 | } | |
278 | } | |
279 | return NULL; | |
280 | } | |
281 | ||
282 | /* | |
283 | * Nothing fancy, just initialize lists and locks and counters. | |
284 | */ | |
285 | void fsnotify_init_mark(struct fsnotify_mark_entry *entry, | |
286 | void (*free_mark)(struct fsnotify_mark_entry *entry)) | |
287 | ||
288 | { | |
289 | spin_lock_init(&entry->lock); | |
290 | atomic_set(&entry->refcnt, 1); | |
291 | INIT_HLIST_NODE(&entry->i_list); | |
292 | entry->group = NULL; | |
293 | entry->mask = 0; | |
294 | entry->inode = NULL; | |
295 | entry->free_mark = free_mark; | |
296 | } | |
297 | ||
298 | /* | |
299 | * Attach an initialized mark entry to a given group and inode. | |
300 | * These marks may be used for the fsnotify backend to determine which | |
301 | * event types should be delivered to which group and for which inodes. | |
302 | */ | |
303 | int fsnotify_add_mark(struct fsnotify_mark_entry *entry, | |
304 | struct fsnotify_group *group, struct inode *inode) | |
305 | { | |
306 | struct fsnotify_mark_entry *lentry; | |
307 | int ret = 0; | |
308 | ||
309 | /* | |
310 | * LOCKING ORDER!!!! | |
311 | * entry->lock | |
312 | * group->mark_lock | |
313 | * inode->i_lock | |
314 | */ | |
315 | spin_lock(&entry->lock); | |
316 | spin_lock(&group->mark_lock); | |
317 | spin_lock(&inode->i_lock); | |
318 | ||
319 | entry->group = group; | |
320 | entry->inode = inode; | |
321 | ||
322 | lentry = fsnotify_find_mark_entry(group, inode); | |
323 | if (!lentry) { | |
324 | hlist_add_head(&entry->i_list, &inode->i_fsnotify_mark_entries); | |
325 | list_add(&entry->g_list, &group->mark_entries); | |
326 | ||
327 | fsnotify_get_mark(entry); /* for i_list and g_list */ | |
328 | ||
329 | atomic_inc(&group->num_marks); | |
330 | ||
331 | fsnotify_recalc_inode_mask_locked(inode); | |
332 | } | |
333 | ||
334 | spin_unlock(&inode->i_lock); | |
335 | spin_unlock(&group->mark_lock); | |
336 | spin_unlock(&entry->lock); | |
337 | ||
338 | if (lentry) { | |
339 | ret = -EEXIST; | |
340 | fsnotify_put_mark(lentry); | |
c28f7e56 EP |
341 | } else { |
342 | __fsnotify_update_child_dentry_flags(inode); | |
3be25f49 EP |
343 | } |
344 | ||
345 | return ret; | |
346 | } |