blob: 3165d85aada2d5ee14a7e46dee6f0c8d07ee8686 [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>
Eric Paris164bc612009-05-21 17:01:58 -040092#include <linux/writeback.h> /* for inode_lock */
Eric Paris3be25f42009-05-21 17:01:26 -040093
94#include <asm/atomic.h>
95
96#include <linux/fsnotify_backend.h>
97#include "fsnotify.h"
98
99void fsnotify_get_mark(struct fsnotify_mark_entry *entry)
100{
101 atomic_inc(&entry->refcnt);
102}
103
104void fsnotify_put_mark(struct fsnotify_mark_entry *entry)
105{
106 if (atomic_dec_and_test(&entry->refcnt))
107 entry->free_mark(entry);
108}
109
110/*
111 * Recalculate the mask of events relevant to a given inode locked.
112 */
113static void fsnotify_recalc_inode_mask_locked(struct inode *inode)
114{
115 struct fsnotify_mark_entry *entry;
116 struct hlist_node *pos;
117 __u32 new_mask = 0;
118
119 assert_spin_locked(&inode->i_lock);
120
121 hlist_for_each_entry(entry, pos, &inode->i_fsnotify_mark_entries, i_list)
122 new_mask |= entry->mask;
123 inode->i_fsnotify_mask = new_mask;
124}
125
126/*
127 * Recalculate the inode->i_fsnotify_mask, or the mask of all FS_* event types
128 * any notifier is interested in hearing for this inode.
129 */
130void fsnotify_recalc_inode_mask(struct inode *inode)
131{
132 spin_lock(&inode->i_lock);
133 fsnotify_recalc_inode_mask_locked(inode);
134 spin_unlock(&inode->i_lock);
Eric Parisc28f7e52009-05-21 17:01:29 -0400135
136 __fsnotify_update_child_dentry_flags(inode);
Eric Paris3be25f42009-05-21 17:01:26 -0400137}
138
139/*
140 * Any time a mark is getting freed we end up here.
141 * The caller had better be holding a reference to this mark so we don't actually
142 * do the final put under the entry->lock
143 */
144void fsnotify_destroy_mark_by_entry(struct fsnotify_mark_entry *entry)
145{
146 struct fsnotify_group *group;
147 struct inode *inode;
148
149 spin_lock(&entry->lock);
150
151 group = entry->group;
152 inode = entry->inode;
153
154 BUG_ON(group && !inode);
155 BUG_ON(!group && inode);
156
157 /* if !group something else already marked this to die */
158 if (!group) {
159 spin_unlock(&entry->lock);
160 return;
161 }
162
163 /* 1 from caller and 1 for being on i_list/g_list */
164 BUG_ON(atomic_read(&entry->refcnt) < 2);
165
166 spin_lock(&group->mark_lock);
167 spin_lock(&inode->i_lock);
168
169 hlist_del_init(&entry->i_list);
170 entry->inode = NULL;
171
172 list_del_init(&entry->g_list);
173 entry->group = NULL;
174
175 fsnotify_put_mark(entry); /* for i_list and g_list */
176
177 /*
178 * this mark is now off the inode->i_fsnotify_mark_entries list and we
179 * hold the inode->i_lock, so this is the perfect time to update the
180 * inode->i_fsnotify_mask
181 */
182 fsnotify_recalc_inode_mask_locked(inode);
183
184 spin_unlock(&inode->i_lock);
185 spin_unlock(&group->mark_lock);
186 spin_unlock(&entry->lock);
187
188 /*
189 * Some groups like to know that marks are being freed. This is a
190 * callback to the group function to let it know that this entry
191 * is being freed.
192 */
Eric Parisa092ee22009-06-11 11:09:48 -0400193 if (group->ops->freeing_mark)
194 group->ops->freeing_mark(entry, group);
Eric Paris3be25f42009-05-21 17:01:26 -0400195
196 /*
Eric Parisc28f7e52009-05-21 17:01:29 -0400197 * __fsnotify_update_child_dentry_flags(inode);
198 *
199 * I really want to call that, but we can't, we have no idea if the inode
200 * still exists the second we drop the entry->lock.
201 *
202 * The next time an event arrive to this inode from one of it's children
203 * __fsnotify_parent will see that the inode doesn't care about it's
204 * children and will update all of these flags then. So really this
205 * is just a lazy update (and could be a perf win...)
206 */
207
208
Eric Paris1ef5f132009-05-21 17:01:54 -0400209 iput(inode);
210
Eric Parisc28f7e52009-05-21 17:01:29 -0400211 /*
Eric Paris3be25f42009-05-21 17:01:26 -0400212 * it's possible that this group tried to destroy itself, but this
213 * this mark was simultaneously being freed by inode. If that's the
214 * case, we finish freeing the group here.
215 */
216 if (unlikely(atomic_dec_and_test(&group->num_marks)))
217 fsnotify_final_destroy_group(group);
218}
219
220/*
221 * Given a group, destroy all of the marks associated with that group.
222 */
223void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
224{
225 struct fsnotify_mark_entry *lentry, *entry;
226 LIST_HEAD(free_list);
227
228 spin_lock(&group->mark_lock);
229 list_for_each_entry_safe(entry, lentry, &group->mark_entries, g_list) {
230 list_add(&entry->free_g_list, &free_list);
231 list_del_init(&entry->g_list);
232 fsnotify_get_mark(entry);
233 }
234 spin_unlock(&group->mark_lock);
235
236 list_for_each_entry_safe(entry, lentry, &free_list, free_g_list) {
237 fsnotify_destroy_mark_by_entry(entry);
238 fsnotify_put_mark(entry);
239 }
240}
241
242/*
243 * Given an inode, destroy all of the marks associated with that inode.
244 */
245void fsnotify_clear_marks_by_inode(struct inode *inode)
246{
247 struct fsnotify_mark_entry *entry, *lentry;
248 struct hlist_node *pos, *n;
249 LIST_HEAD(free_list);
250
251 spin_lock(&inode->i_lock);
252 hlist_for_each_entry_safe(entry, pos, n, &inode->i_fsnotify_mark_entries, i_list) {
253 list_add(&entry->free_i_list, &free_list);
254 hlist_del_init(&entry->i_list);
255 fsnotify_get_mark(entry);
256 }
257 spin_unlock(&inode->i_lock);
258
259 list_for_each_entry_safe(entry, lentry, &free_list, free_i_list) {
260 fsnotify_destroy_mark_by_entry(entry);
261 fsnotify_put_mark(entry);
262 }
263}
264
265/*
266 * given a group and inode, find the mark associated with that combination.
267 * if found take a reference to that mark and return it, else return NULL
268 */
269struct fsnotify_mark_entry *fsnotify_find_mark_entry(struct fsnotify_group *group,
270 struct inode *inode)
271{
272 struct fsnotify_mark_entry *entry;
273 struct hlist_node *pos;
274
275 assert_spin_locked(&inode->i_lock);
276
277 hlist_for_each_entry(entry, pos, &inode->i_fsnotify_mark_entries, i_list) {
278 if (entry->group == group) {
279 fsnotify_get_mark(entry);
280 return entry;
281 }
282 }
283 return NULL;
284}
285
286/*
287 * Nothing fancy, just initialize lists and locks and counters.
288 */
289void fsnotify_init_mark(struct fsnotify_mark_entry *entry,
290 void (*free_mark)(struct fsnotify_mark_entry *entry))
291
292{
293 spin_lock_init(&entry->lock);
294 atomic_set(&entry->refcnt, 1);
295 INIT_HLIST_NODE(&entry->i_list);
296 entry->group = NULL;
297 entry->mask = 0;
298 entry->inode = NULL;
299 entry->free_mark = free_mark;
300}
301
302/*
303 * Attach an initialized mark entry to a given group and inode.
304 * These marks may be used for the fsnotify backend to determine which
305 * event types should be delivered to which group and for which inodes.
306 */
307int fsnotify_add_mark(struct fsnotify_mark_entry *entry,
308 struct fsnotify_group *group, struct inode *inode)
309{
310 struct fsnotify_mark_entry *lentry;
311 int ret = 0;
312
Eric Paris1ef5f132009-05-21 17:01:54 -0400313 inode = igrab(inode);
314 if (unlikely(!inode))
315 return -EINVAL;
316
Eric Paris3be25f42009-05-21 17:01:26 -0400317 /*
318 * LOCKING ORDER!!!!
319 * entry->lock
320 * group->mark_lock
321 * inode->i_lock
322 */
323 spin_lock(&entry->lock);
324 spin_lock(&group->mark_lock);
325 spin_lock(&inode->i_lock);
326
Eric Paris3be25f42009-05-21 17:01:26 -0400327 lentry = fsnotify_find_mark_entry(group, inode);
328 if (!lentry) {
Eric Paris9f0d7932009-09-11 13:03:19 -0400329 entry->group = group;
330 entry->inode = inode;
331
Eric Paris3be25f42009-05-21 17:01:26 -0400332 hlist_add_head(&entry->i_list, &inode->i_fsnotify_mark_entries);
333 list_add(&entry->g_list, &group->mark_entries);
334
335 fsnotify_get_mark(entry); /* for i_list and g_list */
336
337 atomic_inc(&group->num_marks);
338
339 fsnotify_recalc_inode_mask_locked(inode);
340 }
341
342 spin_unlock(&inode->i_lock);
343 spin_unlock(&group->mark_lock);
344 spin_unlock(&entry->lock);
345
346 if (lentry) {
347 ret = -EEXIST;
Eric Paris1ef5f132009-05-21 17:01:54 -0400348 iput(inode);
Eric Paris3be25f42009-05-21 17:01:26 -0400349 fsnotify_put_mark(lentry);
Eric Parisc28f7e52009-05-21 17:01:29 -0400350 } else {
351 __fsnotify_update_child_dentry_flags(inode);
Eric Paris3be25f42009-05-21 17:01:26 -0400352 }
353
354 return ret;
355}
Eric Paris164bc612009-05-21 17:01:58 -0400356
357/**
358 * fsnotify_unmount_inodes - an sb is unmounting. handle any watched inodes.
359 * @list: list of inodes being unmounted (sb->s_inodes)
360 *
361 * Called with inode_lock held, protecting the unmounting super block's list
362 * of inodes, and with iprune_mutex held, keeping shrink_icache_memory() at bay.
363 * We temporarily drop inode_lock, however, and CAN block.
364 */
365void fsnotify_unmount_inodes(struct list_head *list)
366{
367 struct inode *inode, *next_i, *need_iput = NULL;
368
369 list_for_each_entry_safe(inode, next_i, list, i_sb_list) {
370 struct inode *need_iput_tmp;
371
372 /*
373 * We cannot __iget() an inode in state I_CLEAR, I_FREEING,
374 * I_WILL_FREE, or I_NEW which is fine because by that point
375 * the inode cannot have any associated watches.
376 */
377 if (inode->i_state & (I_CLEAR|I_FREEING|I_WILL_FREE|I_NEW))
378 continue;
379
380 /*
381 * If i_count is zero, the inode cannot have any watches and
382 * doing an __iget/iput with MS_ACTIVE clear would actually
383 * evict all inodes with zero i_count from icache which is
384 * unnecessarily violent and may in fact be illegal to do.
385 */
386 if (!atomic_read(&inode->i_count))
387 continue;
388
389 need_iput_tmp = need_iput;
390 need_iput = NULL;
391
392 /* In case fsnotify_inode_delete() drops a reference. */
393 if (inode != need_iput_tmp)
394 __iget(inode);
395 else
396 need_iput_tmp = NULL;
397
398 /* In case the dropping of a reference would nuke next_i. */
399 if ((&next_i->i_sb_list != list) &&
400 atomic_read(&next_i->i_count) &&
401 !(next_i->i_state & (I_CLEAR | I_FREEING | I_WILL_FREE))) {
402 __iget(next_i);
403 need_iput = next_i;
404 }
405
406 /*
407 * We can safely drop inode_lock here because we hold
408 * references on both inode and next_i. Also no new inodes
409 * will be added since the umount has begun. Finally,
410 * iprune_mutex keeps shrink_icache_memory() away.
411 */
412 spin_unlock(&inode_lock);
413
414 if (need_iput_tmp)
415 iput(need_iput_tmp);
416
417 /* for each watch, send FS_UNMOUNT and then remove it */
418 fsnotify(inode, FS_UNMOUNT, inode, FSNOTIFY_EVENT_INODE, NULL, 0);
419
420 fsnotify_inode_delete(inode);
421
422 iput(inode);
423
424 spin_lock(&inode_lock);
425 }
426}