xref: /linux/fs/notify/mark.c (revision 75c1be487a690db43da2c1234fcacd84c982803c)
15444e298SEric Paris /*
25444e298SEric Paris  *  Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
35444e298SEric Paris  *
45444e298SEric Paris  *  This program is free software; you can redistribute it and/or modify
55444e298SEric Paris  *  it under the terms of the GNU General Public License as published by
65444e298SEric Paris  *  the Free Software Foundation; either version 2, or (at your option)
75444e298SEric Paris  *  any later version.
85444e298SEric Paris  *
95444e298SEric Paris  *  This program is distributed in the hope that it will be useful,
105444e298SEric Paris  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
115444e298SEric Paris  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
125444e298SEric Paris  *  GNU General Public License for more details.
135444e298SEric Paris  *
145444e298SEric Paris  *  You should have received a copy of the GNU General Public License
155444e298SEric Paris  *  along with this program; see the file COPYING.  If not, write to
165444e298SEric Paris  *  the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
175444e298SEric Paris  */
185444e298SEric Paris 
195444e298SEric Paris /*
205444e298SEric Paris  * fsnotify inode mark locking/lifetime/and refcnting
215444e298SEric Paris  *
225444e298SEric Paris  * REFCNT:
235444e298SEric Paris  * The mark->refcnt tells how many "things" in the kernel currently are
245444e298SEric Paris  * referencing this object.  The object typically will live inside the kernel
255444e298SEric Paris  * with a refcnt of 2, one for each list it is on (i_list, g_list).  Any task
265444e298SEric Paris  * which can find this object holding the appropriete locks, can take a reference
275444e298SEric Paris  * and the object itself is guarenteed to survive until the reference is dropped.
285444e298SEric Paris  *
295444e298SEric Paris  * LOCKING:
305444e298SEric Paris  * There are 3 spinlocks involved with fsnotify inode marks and they MUST
315444e298SEric Paris  * be taken in order as follows:
325444e298SEric Paris  *
335444e298SEric Paris  * mark->lock
345444e298SEric Paris  * group->mark_lock
355444e298SEric Paris  * inode->i_lock
365444e298SEric Paris  *
375444e298SEric Paris  * mark->lock protects 2 things, mark->group and mark->inode.  You must hold
385444e298SEric Paris  * that lock to dereference either of these things (they could be NULL even with
395444e298SEric Paris  * the lock)
405444e298SEric Paris  *
415444e298SEric Paris  * group->mark_lock protects the marks_list anchored inside a given group
425444e298SEric Paris  * and each mark is hooked via the g_list.  It also sorta protects the
435444e298SEric Paris  * free_g_list, which when used is anchored by a private list on the stack of the
445444e298SEric Paris  * task which held the group->mark_lock.
455444e298SEric Paris  *
465444e298SEric Paris  * inode->i_lock protects the i_fsnotify_marks list anchored inside a
475444e298SEric Paris  * given inode and each mark is hooked via the i_list. (and sorta the
485444e298SEric Paris  * free_i_list)
495444e298SEric Paris  *
505444e298SEric Paris  *
515444e298SEric Paris  * LIFETIME:
525444e298SEric Paris  * Inode marks survive between when they are added to an inode and when their
535444e298SEric Paris  * refcnt==0.
545444e298SEric Paris  *
555444e298SEric Paris  * The inode mark can be cleared for a number of different reasons including:
565444e298SEric Paris  * - The inode is unlinked for the last time.  (fsnotify_inode_remove)
575444e298SEric Paris  * - The inode is being evicted from cache. (fsnotify_inode_delete)
585444e298SEric Paris  * - The fs the inode is on is unmounted.  (fsnotify_inode_delete/fsnotify_unmount_inodes)
595444e298SEric Paris  * - Something explicitly requests that it be removed.  (fsnotify_destroy_mark)
605444e298SEric Paris  * - The fsnotify_group associated with the mark is going away and all such marks
615444e298SEric Paris  *   need to be cleaned up. (fsnotify_clear_marks_by_group)
625444e298SEric Paris  *
635444e298SEric Paris  * Worst case we are given an inode and need to clean up all the marks on that
645444e298SEric Paris  * inode.  We take i_lock and walk the i_fsnotify_marks safely.  For each
655444e298SEric Paris  * mark on the list we take a reference (so the mark can't disappear under us).
665444e298SEric Paris  * We remove that mark form the inode's list of marks and we add this mark to a
675444e298SEric Paris  * private list anchored on the stack using i_free_list;  At this point we no
685444e298SEric Paris  * longer fear anything finding the mark using the inode's list of marks.
695444e298SEric Paris  *
705444e298SEric Paris  * We can safely and locklessly run the private list on the stack of everything
715444e298SEric Paris  * we just unattached from the original inode.  For each mark on the private list
725444e298SEric Paris  * we grab the mark-> and can thus dereference mark->group and mark->inode.  If
735444e298SEric Paris  * we see the group and inode are not NULL we take those locks.  Now holding all
745444e298SEric Paris  * 3 locks we can completely remove the mark from other tasks finding it in the
755444e298SEric Paris  * future.  Remember, 10 things might already be referencing this mark, but they
765444e298SEric Paris  * better be holding a ref.  We drop our reference we took before we unhooked it
775444e298SEric Paris  * from the inode.  When the ref hits 0 we can free the mark.
785444e298SEric Paris  *
795444e298SEric Paris  * Very similarly for freeing by group, except we use free_g_list.
805444e298SEric Paris  *
815444e298SEric Paris  * This has the very interesting property of being able to run concurrently with
825444e298SEric Paris  * any (or all) other directions.
835444e298SEric Paris  */
845444e298SEric Paris 
855444e298SEric Paris #include <linux/fs.h>
865444e298SEric Paris #include <linux/init.h>
875444e298SEric Paris #include <linux/kernel.h>
88*75c1be48SEric Paris #include <linux/kthread.h>
895444e298SEric Paris #include <linux/module.h>
905444e298SEric Paris #include <linux/mutex.h>
915444e298SEric Paris #include <linux/slab.h>
925444e298SEric Paris #include <linux/spinlock.h>
93*75c1be48SEric Paris #include <linux/srcu.h>
945444e298SEric Paris #include <linux/writeback.h> /* for inode_lock */
955444e298SEric Paris 
965444e298SEric Paris #include <asm/atomic.h>
975444e298SEric Paris 
985444e298SEric Paris #include <linux/fsnotify_backend.h>
995444e298SEric Paris #include "fsnotify.h"
1005444e298SEric Paris 
101*75c1be48SEric Paris struct srcu_struct fsnotify_mark_srcu;
102*75c1be48SEric Paris static DEFINE_SPINLOCK(destroy_lock);
103*75c1be48SEric Paris static LIST_HEAD(destroy_list);
104*75c1be48SEric Paris static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
105*75c1be48SEric Paris 
1065444e298SEric Paris void fsnotify_get_mark(struct fsnotify_mark *mark)
1075444e298SEric Paris {
1085444e298SEric Paris 	atomic_inc(&mark->refcnt);
1095444e298SEric Paris }
1105444e298SEric Paris 
1115444e298SEric Paris void fsnotify_put_mark(struct fsnotify_mark *mark)
1125444e298SEric Paris {
1135444e298SEric Paris 	if (atomic_dec_and_test(&mark->refcnt))
1145444e298SEric Paris 		mark->free_mark(mark);
1155444e298SEric Paris }
1165444e298SEric Paris 
1175444e298SEric Paris /*
1185444e298SEric Paris  * Any time a mark is getting freed we end up here.
1195444e298SEric Paris  * The caller had better be holding a reference to this mark so we don't actually
1205444e298SEric Paris  * do the final put under the mark->lock
1215444e298SEric Paris  */
1225444e298SEric Paris void fsnotify_destroy_mark(struct fsnotify_mark *mark)
1235444e298SEric Paris {
1245444e298SEric Paris 	struct fsnotify_group *group;
1250d48b7f0SEric Paris 	struct inode *inode = NULL;
1265444e298SEric Paris 
1275444e298SEric Paris 	spin_lock(&mark->lock);
1285444e298SEric Paris 
1295444e298SEric Paris 	group = mark->group;
1305444e298SEric Paris 
131700307a2SEric Paris 	/* something else already called this function on this mark */
132700307a2SEric Paris 	if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
1335444e298SEric Paris 		spin_unlock(&mark->lock);
1345444e298SEric Paris 		return;
1355444e298SEric Paris 	}
1365444e298SEric Paris 
137700307a2SEric Paris 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
138700307a2SEric Paris 
1395444e298SEric Paris 	/* 1 from caller and 1 for being on i_list/g_list */
1405444e298SEric Paris 	BUG_ON(atomic_read(&mark->refcnt) < 2);
1415444e298SEric Paris 
1425444e298SEric Paris 	spin_lock(&group->mark_lock);
1435444e298SEric Paris 
1440d48b7f0SEric Paris 	if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
1450d48b7f0SEric Paris 		inode = mark->i.inode;
146b31d397eSEric Paris 		fsnotify_destroy_inode_mark(mark);
1470d48b7f0SEric Paris 	} else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
1480d48b7f0SEric Paris 		fsnotify_destroy_vfsmount_mark(mark);
1495444e298SEric Paris 	else
1505444e298SEric Paris 		BUG();
1515444e298SEric Paris 
1525444e298SEric Paris 	list_del_init(&mark->g_list);
1535444e298SEric Paris 
1545444e298SEric Paris 	spin_unlock(&group->mark_lock);
1555444e298SEric Paris 	spin_unlock(&mark->lock);
1565444e298SEric Paris 
157*75c1be48SEric Paris 	spin_lock(&destroy_lock);
158*75c1be48SEric Paris 	list_add(&mark->destroy_list, &destroy_list);
159*75c1be48SEric Paris 	spin_unlock(&destroy_lock);
160*75c1be48SEric Paris 	wake_up(&destroy_waitq);
161*75c1be48SEric Paris 
1625444e298SEric Paris 	/*
1635444e298SEric Paris 	 * Some groups like to know that marks are being freed.  This is a
1645444e298SEric Paris 	 * callback to the group function to let it know that this mark
1655444e298SEric Paris 	 * is being freed.
1665444e298SEric Paris 	 */
1675444e298SEric Paris 	if (group->ops->freeing_mark)
1685444e298SEric Paris 		group->ops->freeing_mark(mark, group);
1695444e298SEric Paris 
1705444e298SEric Paris 	/*
1715444e298SEric Paris 	 * __fsnotify_update_child_dentry_flags(inode);
1725444e298SEric Paris 	 *
1735444e298SEric Paris 	 * I really want to call that, but we can't, we have no idea if the inode
1745444e298SEric Paris 	 * still exists the second we drop the mark->lock.
1755444e298SEric Paris 	 *
1765444e298SEric Paris 	 * The next time an event arrive to this inode from one of it's children
1775444e298SEric Paris 	 * __fsnotify_parent will see that the inode doesn't care about it's
1785444e298SEric Paris 	 * children and will update all of these flags then.  So really this
1795444e298SEric Paris 	 * is just a lazy update (and could be a perf win...)
1805444e298SEric Paris 	 */
1815444e298SEric Paris 
18290b1e7a5SEric Paris 	if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
1835444e298SEric Paris 		iput(inode);
1845444e298SEric Paris 
1855444e298SEric Paris 	/*
1865444e298SEric Paris 	 * it's possible that this group tried to destroy itself, but this
1875444e298SEric Paris 	 * this mark was simultaneously being freed by inode.  If that's the
1885444e298SEric Paris 	 * case, we finish freeing the group here.
1895444e298SEric Paris 	 */
1905444e298SEric Paris 	if (unlikely(atomic_dec_and_test(&group->num_marks)))
1915444e298SEric Paris 		fsnotify_final_destroy_group(group);
1925444e298SEric Paris }
1935444e298SEric Paris 
19490b1e7a5SEric Paris void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
19590b1e7a5SEric Paris {
19690b1e7a5SEric Paris 	assert_spin_locked(&mark->lock);
19790b1e7a5SEric Paris 
19890b1e7a5SEric Paris 	mark->mask = mask;
19990b1e7a5SEric Paris 
20090b1e7a5SEric Paris 	if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
20190b1e7a5SEric Paris 		fsnotify_set_inode_mark_mask_locked(mark, mask);
20290b1e7a5SEric Paris }
20390b1e7a5SEric Paris 
20433af5e32SEric Paris void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
20533af5e32SEric Paris {
20633af5e32SEric Paris 	assert_spin_locked(&mark->lock);
20733af5e32SEric Paris 
20833af5e32SEric Paris 	mark->ignored_mask = mask;
20933af5e32SEric Paris }
21090b1e7a5SEric Paris 
2115444e298SEric Paris /*
2125444e298SEric Paris  * Attach an initialized mark to a given group and fs object.
2135444e298SEric Paris  * These marks may be used for the fsnotify backend to determine which
2145444e298SEric Paris  * event types should be delivered to which group.
2155444e298SEric Paris  */
2165444e298SEric Paris int fsnotify_add_mark(struct fsnotify_mark *mark,
2175444e298SEric Paris 		      struct fsnotify_group *group, struct inode *inode,
2185444e298SEric Paris 		      struct vfsmount *mnt, int allow_dups)
2195444e298SEric Paris {
2205444e298SEric Paris 	int ret = 0;
2215444e298SEric Paris 
2225444e298SEric Paris 	BUG_ON(inode && mnt);
2235444e298SEric Paris 	BUG_ON(!inode && !mnt);
2245444e298SEric Paris 
2255444e298SEric Paris 	/*
2265444e298SEric Paris 	 * if this group isn't being testing for inode type events we need
2275444e298SEric Paris 	 * to start testing
2285444e298SEric Paris 	 */
2295444e298SEric Paris 	if (inode && unlikely(list_empty(&group->inode_group_list)))
2305444e298SEric Paris 		fsnotify_add_inode_group(group);
2315444e298SEric Paris 	else if (mnt && unlikely(list_empty(&group->vfsmount_group_list)))
2325444e298SEric Paris 		fsnotify_add_vfsmount_group(group);
2335444e298SEric Paris 
2345444e298SEric Paris 	/*
2355444e298SEric Paris 	 * LOCKING ORDER!!!!
2365444e298SEric Paris 	 * mark->lock
2375444e298SEric Paris 	 * group->mark_lock
2385444e298SEric Paris 	 * inode->i_lock
2395444e298SEric Paris 	 */
2405444e298SEric Paris 	spin_lock(&mark->lock);
2415444e298SEric Paris 	spin_lock(&group->mark_lock);
2425444e298SEric Paris 
243700307a2SEric Paris 	mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE;
244700307a2SEric Paris 
2455444e298SEric Paris 	mark->group = group;
2465444e298SEric Paris 	list_add(&mark->g_list, &group->marks_list);
2475444e298SEric Paris 	atomic_inc(&group->num_marks);
2485444e298SEric Paris 	fsnotify_get_mark(mark); /* for i_list and g_list */
2495444e298SEric Paris 
2505444e298SEric Paris 	if (inode) {
2515444e298SEric Paris 		ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
2525444e298SEric Paris 		if (ret)
2535444e298SEric Paris 			goto err;
2540d48b7f0SEric Paris 	} else if (mnt) {
2550d48b7f0SEric Paris 		ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
2560d48b7f0SEric Paris 		if (ret)
2570d48b7f0SEric Paris 			goto err;
2585444e298SEric Paris 	} else {
2595444e298SEric Paris 		BUG();
2605444e298SEric Paris 	}
2615444e298SEric Paris 
2625444e298SEric Paris 	spin_unlock(&group->mark_lock);
26390b1e7a5SEric Paris 
26490b1e7a5SEric Paris 	/* this will pin the object if appropriate */
26590b1e7a5SEric Paris 	fsnotify_set_mark_mask_locked(mark, mark->mask);
26690b1e7a5SEric Paris 
2675444e298SEric Paris 	spin_unlock(&mark->lock);
2685444e298SEric Paris 
2695444e298SEric Paris 	if (inode)
2705444e298SEric Paris 		__fsnotify_update_child_dentry_flags(inode);
2715444e298SEric Paris 
2725444e298SEric Paris 	return ret;
2735444e298SEric Paris err:
274700307a2SEric Paris 	mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
2755444e298SEric Paris 	list_del_init(&mark->g_list);
276*75c1be48SEric Paris 	mark->group = NULL;
2775444e298SEric Paris 	atomic_dec(&group->num_marks);
2785444e298SEric Paris 
2795444e298SEric Paris 	spin_unlock(&group->mark_lock);
2805444e298SEric Paris 	spin_unlock(&mark->lock);
2815444e298SEric Paris 
282*75c1be48SEric Paris 	spin_lock(&destroy_lock);
283*75c1be48SEric Paris 	list_add(&mark->destroy_list, &destroy_list);
284*75c1be48SEric Paris 	spin_unlock(&destroy_lock);
285*75c1be48SEric Paris 	wake_up(&destroy_waitq);
286*75c1be48SEric Paris 
2875444e298SEric Paris 	return ret;
2885444e298SEric Paris }
2895444e298SEric Paris 
2905444e298SEric Paris /*
2914d92604cSEric Paris  * clear any marks in a group in which mark->flags & flags is true
2925444e298SEric Paris  */
2934d92604cSEric Paris void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
2944d92604cSEric Paris 					 unsigned int flags)
2955444e298SEric Paris {
2965444e298SEric Paris 	struct fsnotify_mark *lmark, *mark;
2975444e298SEric Paris 	LIST_HEAD(free_list);
2985444e298SEric Paris 
2995444e298SEric Paris 	spin_lock(&group->mark_lock);
3005444e298SEric Paris 	list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
3014d92604cSEric Paris 		if (mark->flags & flags) {
3025444e298SEric Paris 			list_add(&mark->free_g_list, &free_list);
3035444e298SEric Paris 			list_del_init(&mark->g_list);
3045444e298SEric Paris 			fsnotify_get_mark(mark);
3055444e298SEric Paris 		}
3064d92604cSEric Paris 	}
3075444e298SEric Paris 	spin_unlock(&group->mark_lock);
3085444e298SEric Paris 
3095444e298SEric Paris 	list_for_each_entry_safe(mark, lmark, &free_list, free_g_list) {
3105444e298SEric Paris 		fsnotify_destroy_mark(mark);
3115444e298SEric Paris 		fsnotify_put_mark(mark);
3125444e298SEric Paris 	}
3135444e298SEric Paris }
3145444e298SEric Paris 
3154d92604cSEric Paris /*
3164d92604cSEric Paris  * Given a group, destroy all of the marks associated with that group.
3174d92604cSEric Paris  */
3184d92604cSEric Paris void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
3194d92604cSEric Paris {
3204d92604cSEric Paris 	fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
3214d92604cSEric Paris }
3224d92604cSEric Paris 
3235444e298SEric Paris void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
3245444e298SEric Paris {
3255444e298SEric Paris 	assert_spin_locked(&old->lock);
3265444e298SEric Paris 	new->i.inode = old->i.inode;
3275444e298SEric Paris 	new->m.mnt = old->m.mnt;
3285444e298SEric Paris 	new->group = old->group;
3295444e298SEric Paris 	new->mask = old->mask;
3305444e298SEric Paris 	new->free_mark = old->free_mark;
3315444e298SEric Paris }
3325444e298SEric Paris 
3335444e298SEric Paris /*
3345444e298SEric Paris  * Nothing fancy, just initialize lists and locks and counters.
3355444e298SEric Paris  */
3365444e298SEric Paris void fsnotify_init_mark(struct fsnotify_mark *mark,
3375444e298SEric Paris 			void (*free_mark)(struct fsnotify_mark *mark))
3385444e298SEric Paris {
339ba643f04SEric Paris 	memset(mark, 0, sizeof(*mark));
3405444e298SEric Paris 	spin_lock_init(&mark->lock);
3415444e298SEric Paris 	atomic_set(&mark->refcnt, 1);
3425444e298SEric Paris 	mark->free_mark = free_mark;
3435444e298SEric Paris }
344*75c1be48SEric Paris 
345*75c1be48SEric Paris static int fsnotify_mark_destroy(void *ignored)
346*75c1be48SEric Paris {
347*75c1be48SEric Paris 	struct fsnotify_mark *mark, *next;
348*75c1be48SEric Paris 	LIST_HEAD(private_destroy_list);
349*75c1be48SEric Paris 
350*75c1be48SEric Paris 	for (;;) {
351*75c1be48SEric Paris 		spin_lock(&destroy_lock);
352*75c1be48SEric Paris 		list_for_each_entry_safe(mark, next, &destroy_list, destroy_list) {
353*75c1be48SEric Paris 			list_del(&mark->destroy_list);
354*75c1be48SEric Paris 			list_add(&mark->destroy_list, &private_destroy_list);
355*75c1be48SEric Paris 		}
356*75c1be48SEric Paris 		spin_unlock(&destroy_lock);
357*75c1be48SEric Paris 
358*75c1be48SEric Paris 		synchronize_srcu(&fsnotify_mark_srcu);
359*75c1be48SEric Paris 
360*75c1be48SEric Paris 		list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) {
361*75c1be48SEric Paris 			list_del_init(&mark->destroy_list);
362*75c1be48SEric Paris 			fsnotify_put_mark(mark);
363*75c1be48SEric Paris 		}
364*75c1be48SEric Paris 
365*75c1be48SEric Paris 		wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
366*75c1be48SEric Paris 	}
367*75c1be48SEric Paris 
368*75c1be48SEric Paris 	return 0;
369*75c1be48SEric Paris }
370*75c1be48SEric Paris 
371*75c1be48SEric Paris static int __init fsnotify_mark_init(void)
372*75c1be48SEric Paris {
373*75c1be48SEric Paris 	struct task_struct *thread;
374*75c1be48SEric Paris 
375*75c1be48SEric Paris 	thread = kthread_run(fsnotify_mark_destroy, NULL,
376*75c1be48SEric Paris 			     "fsnotify_mark");
377*75c1be48SEric Paris 	if (IS_ERR(thread))
378*75c1be48SEric Paris 		panic("unable to start fsnotify mark destruction thread.");
379*75c1be48SEric Paris 
380*75c1be48SEric Paris 	return 0;
381*75c1be48SEric Paris }
382*75c1be48SEric Paris device_initcall(fsnotify_mark_init);
383