2 * fs/inotify_user.c - inotify support for userspace
5 * John McCutchan <ttb@tentacle.dhs.org>
6 * Robert Love <rml@novell.com>
8 * Copyright (C) 2005 John McCutchan
9 * Copyright 2006 Hewlett-Packard Development Company, L.P.
11 * Copyright (C) 2009 Eric Paris <Red Hat Inc>
12 * inotify was largely rewriten to make use of the fsnotify infrastructure
14 * This program is free software; you can redistribute it and/or modify it
15 * under the terms of the GNU General Public License as published by the
16 * Free Software Foundation; either version 2, or (at your option) any
19 * This program is distributed in the hope that it will be useful, but
20 * WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 * General Public License for more details.
25 #include <linux/file.h>
26 #include <linux/fs.h> /* struct inode */
27 #include <linux/fsnotify_backend.h>
28 #include <linux/idr.h>
29 #include <linux/init.h> /* module_init */
30 #include <linux/inotify.h>
31 #include <linux/kernel.h> /* roundup() */
32 #include <linux/namei.h> /* LOOKUP_FOLLOW */
33 #include <linux/sched.h> /* struct user */
34 #include <linux/slab.h> /* struct kmem_cache */
35 #include <linux/syscalls.h>
36 #include <linux/types.h>
37 #include <linux/anon_inodes.h>
38 #include <linux/uaccess.h>
39 #include <linux/poll.h>
40 #include <linux/wait.h>
44 #include <asm/ioctls.h>
46 /* these are configurable via /proc/sys/fs/inotify/ */
47 static int inotify_max_user_instances __read_mostly;
48 static int inotify_max_queued_events __read_mostly;
49 int inotify_max_user_watches __read_mostly;
51 static struct kmem_cache *inotify_inode_mark_cachep __read_mostly;
52 struct kmem_cache *event_priv_cachep __read_mostly;
55 * When inotify registers a new group it increments this and uses that
56 * value as an offset to set the fsnotify group "name" and priority.
58 static atomic_t inotify_grp_num;
62 #include <linux/sysctl.h>
66 ctl_table inotify_table[] = {
68 .procname = "max_user_instances",
69 .data = &inotify_max_user_instances,
70 .maxlen = sizeof(int),
72 .proc_handler = proc_dointvec_minmax,
76 .procname = "max_user_watches",
77 .data = &inotify_max_user_watches,
78 .maxlen = sizeof(int),
80 .proc_handler = proc_dointvec_minmax,
84 .procname = "max_queued_events",
85 .data = &inotify_max_queued_events,
86 .maxlen = sizeof(int),
88 .proc_handler = proc_dointvec_minmax,
93 #endif /* CONFIG_SYSCTL */
95 static inline __u32 inotify_arg_to_mask(u32 arg)
99 /* everything should accept their own ignored and cares about children */
100 mask = (FS_IN_IGNORED | FS_EVENT_ON_CHILD);
102 /* mask off the flags used to open the fd */
103 mask |= (arg & (IN_ALL_EVENTS | IN_ONESHOT));
108 static inline u32 inotify_mask_to_arg(__u32 mask)
110 return mask & (IN_ALL_EVENTS | IN_ISDIR | IN_UNMOUNT | IN_IGNORED |
114 /* intofiy userspace file descriptor functions */
115 static unsigned int inotify_poll(struct file *file, poll_table *wait)
117 struct fsnotify_group *group = file->private_data;
120 poll_wait(file, &group->notification_waitq, wait);
121 mutex_lock(&group->notification_mutex);
122 if (!fsnotify_notify_queue_is_empty(group))
123 ret = POLLIN | POLLRDNORM;
124 mutex_unlock(&group->notification_mutex);
130 * Get an inotify_kernel_event if one exists and is small
131 * enough to fit in "count". Return an error pointer if
134 * Called with the group->notification_mutex held.
136 static struct fsnotify_event *get_one_event(struct fsnotify_group *group,
139 size_t event_size = sizeof(struct inotify_event);
140 struct fsnotify_event *event;
142 if (fsnotify_notify_queue_is_empty(group))
145 event = fsnotify_peek_notify_event(group);
148 event_size += roundup(event->name_len + 1, event_size);
150 if (event_size > count)
151 return ERR_PTR(-EINVAL);
153 /* held the notification_mutex the whole time, so this is the
154 * same event we peeked above */
155 fsnotify_remove_notify_event(group);
161 * Copy an event to user space, returning how much we copied.
163 * We already checked that the event size is smaller than the
164 * buffer we had in "get_one_event()" above.
166 static ssize_t copy_event_to_user(struct fsnotify_group *group,
167 struct fsnotify_event *event,
170 struct inotify_event inotify_event;
171 struct fsnotify_event_private_data *fsn_priv;
172 struct inotify_event_private_data *priv;
173 size_t event_size = sizeof(struct inotify_event);
176 /* we get the inotify watch descriptor from the event private data */
177 spin_lock(&event->lock);
178 fsn_priv = fsnotify_remove_priv_from_event(group, event);
179 spin_unlock(&event->lock);
182 inotify_event.wd = -1;
184 priv = container_of(fsn_priv, struct inotify_event_private_data,
185 fsnotify_event_priv_data);
186 inotify_event.wd = priv->wd;
187 inotify_free_event_priv(fsn_priv);
191 * round up event->name_len so it is a multiple of event_size
192 * plus an extra byte for the terminating '\0'.
195 name_len = roundup(event->name_len + 1, event_size);
196 inotify_event.len = name_len;
198 inotify_event.mask = inotify_mask_to_arg(event->mask);
199 inotify_event.cookie = event->sync_cookie;
201 /* send the main event */
202 if (copy_to_user(buf, &inotify_event, event_size))
208 * fsnotify only stores the pathname, so here we have to send the pathname
209 * and then pad that pathname out to a multiple of sizeof(inotify_event)
210 * with zeros. I get my zeros from the nul_inotify_event.
213 unsigned int len_to_zero = name_len - event->name_len;
214 /* copy the path name */
215 if (copy_to_user(buf, event->file_name, event->name_len))
217 buf += event->name_len;
219 /* fill userspace with 0's */
220 if (clear_user(buf, len_to_zero))
223 event_size += name_len;
229 static ssize_t inotify_read(struct file *file, char __user *buf,
230 size_t count, loff_t *pos)
232 struct fsnotify_group *group;
233 struct fsnotify_event *kevent;
239 group = file->private_data;
242 prepare_to_wait(&group->notification_waitq, &wait, TASK_INTERRUPTIBLE);
244 mutex_lock(&group->notification_mutex);
245 kevent = get_one_event(group, count);
246 mutex_unlock(&group->notification_mutex);
249 ret = PTR_ERR(kevent);
252 ret = copy_event_to_user(group, kevent, buf);
253 fsnotify_put_event(kevent);
262 if (file->f_flags & O_NONBLOCK)
265 if (signal_pending(current))
274 finish_wait(&group->notification_waitq, &wait);
275 if (start != buf && ret != -EFAULT)
280 static int inotify_fasync(int fd, struct file *file, int on)
282 struct fsnotify_group *group = file->private_data;
284 return fasync_helper(fd, file, on, &group->inotify_data.fa) >= 0 ? 0 : -EIO;
287 static int inotify_release(struct inode *ignored, struct file *file)
289 struct fsnotify_group *group = file->private_data;
290 struct user_struct *user = group->inotify_data.user;
292 fsnotify_clear_marks_by_group(group);
294 /* free this group, matching get was inotify_init->fsnotify_obtain_group */
295 fsnotify_put_group(group);
297 atomic_dec(&user->inotify_devs);
302 static long inotify_ioctl(struct file *file, unsigned int cmd,
305 struct fsnotify_group *group;
306 struct fsnotify_event_holder *holder;
307 struct fsnotify_event *event;
312 group = file->private_data;
313 p = (void __user *) arg;
317 mutex_lock(&group->notification_mutex);
318 list_for_each_entry(holder, &group->notification_list, event_list) {
319 event = holder->event;
320 send_len += sizeof(struct inotify_event);
322 send_len += roundup(event->name_len + 1,
323 sizeof(struct inotify_event));
325 mutex_unlock(&group->notification_mutex);
326 ret = put_user(send_len, (int __user *) p);
333 static const struct file_operations inotify_fops = {
334 .poll = inotify_poll,
335 .read = inotify_read,
336 .fasync = inotify_fasync,
337 .release = inotify_release,
338 .unlocked_ioctl = inotify_ioctl,
339 .compat_ioctl = inotify_ioctl,
344 * find_inode - resolve a user-given path to a specific inode
346 static int inotify_find_inode(const char __user *dirname, struct path *path, unsigned flags)
350 error = user_path_at(AT_FDCWD, dirname, flags, path);
353 /* you can only watch an inode if you have read permissions on it */
354 error = inode_permission(path->dentry->d_inode, MAY_READ);
360 static int inotify_add_to_idr(struct idr *idr, spinlock_t *idr_lock,
362 struct inotify_inode_mark_entry *ientry)
367 if (unlikely(!idr_pre_get(idr, GFP_KERNEL)))
371 ret = idr_get_new_above(idr, ientry, *last_wd + 1,
373 /* we added the mark to the idr, take a reference */
375 fsnotify_get_mark(&ientry->fsn_entry);
376 *last_wd = ientry->wd;
378 spin_unlock(idr_lock);
379 } while (ret == -EAGAIN);
384 static struct inotify_inode_mark_entry *inotify_idr_find_locked(struct fsnotify_group *group,
387 struct idr *idr = &group->inotify_data.idr;
388 spinlock_t *idr_lock = &group->inotify_data.idr_lock;
389 struct inotify_inode_mark_entry *ientry;
391 assert_spin_locked(idr_lock);
393 ientry = idr_find(idr, wd);
395 struct fsnotify_mark_entry *fsn_entry = &ientry->fsn_entry;
397 fsnotify_get_mark(fsn_entry);
398 /* One ref for being in the idr, one ref we just took */
399 BUG_ON(atomic_read(&fsn_entry->refcnt) < 2);
405 static struct inotify_inode_mark_entry *inotify_idr_find(struct fsnotify_group *group,
408 struct inotify_inode_mark_entry *ientry;
409 spinlock_t *idr_lock = &group->inotify_data.idr_lock;
412 ientry = inotify_idr_find_locked(group, wd);
413 spin_unlock(idr_lock);
418 static void do_inotify_remove_from_idr(struct fsnotify_group *group,
419 struct inotify_inode_mark_entry *ientry)
421 struct idr *idr = &group->inotify_data.idr;
422 spinlock_t *idr_lock = &group->inotify_data.idr_lock;
425 assert_spin_locked(idr_lock);
429 /* removed from the idr, drop that ref */
430 fsnotify_put_mark(&ientry->fsn_entry);
434 * Remove the mark from the idr (if present) and drop the reference
435 * on the mark because it was in the idr.
437 static void inotify_remove_from_idr(struct fsnotify_group *group,
438 struct inotify_inode_mark_entry *ientry)
440 spinlock_t *idr_lock = &group->inotify_data.idr_lock;
441 struct inotify_inode_mark_entry *found_ientry = NULL;
448 * does this ientry think it is in the idr? we shouldn't get called
452 printk(KERN_WARNING "%s: ientry=%p ientry->wd=%d ientry->group=%p"
453 " ientry->inode=%p\n", __func__, ientry, ientry->wd,
454 ientry->fsn_entry.group, ientry->fsn_entry.inode);
459 /* Lets look in the idr to see if we find it */
460 found_ientry = inotify_idr_find_locked(group, wd);
461 if (unlikely(!found_ientry)) {
462 printk(KERN_WARNING "%s: ientry=%p ientry->wd=%d ientry->group=%p"
463 " ientry->inode=%p\n", __func__, ientry, ientry->wd,
464 ientry->fsn_entry.group, ientry->fsn_entry.inode);
470 * We found an entry in the idr at the right wd, but it's
471 * not the entry we were told to remove. eparis seriously
472 * fucked up somewhere.
474 if (unlikely(found_ientry != ientry)) {
476 printk(KERN_WARNING "%s: ientry=%p ientry->wd=%d ientry->group=%p "
477 "entry->inode=%p found_ientry=%p found_ientry->wd=%d "
478 "found_ientry->group=%p found_ientry->inode=%p\n",
479 __func__, ientry, ientry->wd, ientry->fsn_entry.group,
480 ientry->fsn_entry.inode, found_ientry, found_ientry->wd,
481 found_ientry->fsn_entry.group,
482 found_ientry->fsn_entry.inode);
487 * One ref for being in the idr
488 * one ref held by the caller trying to kill us
489 * one ref grabbed by inotify_idr_find
491 if (unlikely(atomic_read(&ientry->fsn_entry.refcnt) < 3)) {
492 printk(KERN_WARNING "%s: ientry=%p ientry->wd=%d ientry->group=%p"
493 " ientry->inode=%p\n", __func__, ientry, ientry->wd,
494 ientry->fsn_entry.group, ientry->fsn_entry.inode);
495 /* we can't really recover with bad ref cnting.. */
499 do_inotify_remove_from_idr(group, ientry);
501 /* match the ref taken by inotify_idr_find_locked() */
503 fsnotify_put_mark(&found_ientry->fsn_entry);
505 spin_unlock(idr_lock);
509 * Send IN_IGNORED for this wd, remove this wd from the idr.
511 void inotify_ignored_and_remove_idr(struct fsnotify_mark_entry *entry,
512 struct fsnotify_group *group)
514 struct inotify_inode_mark_entry *ientry;
515 struct fsnotify_event *ignored_event;
516 struct inotify_event_private_data *event_priv;
517 struct fsnotify_event_private_data *fsn_event_priv;
520 ignored_event = fsnotify_create_event(NULL, FS_IN_IGNORED, NULL,
521 FSNOTIFY_EVENT_NONE, NULL, 0,
526 ientry = container_of(entry, struct inotify_inode_mark_entry, fsn_entry);
528 event_priv = kmem_cache_alloc(event_priv_cachep, GFP_NOFS);
529 if (unlikely(!event_priv))
530 goto skip_send_ignore;
532 fsn_event_priv = &event_priv->fsnotify_event_priv_data;
534 fsn_event_priv->group = group;
535 event_priv->wd = ientry->wd;
537 ret = fsnotify_add_notify_event(group, ignored_event, fsn_event_priv);
539 inotify_free_event_priv(fsn_event_priv);
543 /* matches the reference taken when the event was created */
544 fsnotify_put_event(ignored_event);
546 /* remove this entry from the idr */
547 inotify_remove_from_idr(group, ientry);
549 atomic_dec(&group->inotify_data.user->inotify_watches);
552 /* ding dong the mark is dead */
553 static void inotify_free_mark(struct fsnotify_mark_entry *entry)
555 struct inotify_inode_mark_entry *ientry;
557 ientry = container_of(entry, struct inotify_inode_mark_entry, fsn_entry);
559 kmem_cache_free(inotify_inode_mark_cachep, ientry);
562 static int inotify_update_existing_watch(struct fsnotify_group *group,
566 struct fsnotify_mark_entry *entry;
567 struct inotify_inode_mark_entry *ientry;
568 __u32 old_mask, new_mask;
570 int add = (arg & IN_MASK_ADD);
573 /* don't allow invalid bits: we don't want flags set */
574 mask = inotify_arg_to_mask(arg);
578 spin_lock(&inode->i_lock);
579 entry = fsnotify_find_mark_entry(group, inode);
580 spin_unlock(&inode->i_lock);
584 ientry = container_of(entry, struct inotify_inode_mark_entry, fsn_entry);
586 spin_lock(&entry->lock);
588 old_mask = entry->mask;
591 new_mask = entry->mask;
594 new_mask = entry->mask;
597 spin_unlock(&entry->lock);
599 if (old_mask != new_mask) {
600 /* more bits in old than in new? */
601 int dropped = (old_mask & ~new_mask);
602 /* more bits in this entry than the inode's mask? */
603 int do_inode = (new_mask & ~inode->i_fsnotify_mask);
604 /* more bits in this entry than the group? */
605 int do_group = (new_mask & ~group->mask);
607 /* update the inode with this new entry */
608 if (dropped || do_inode)
609 fsnotify_recalc_inode_mask(inode);
611 /* update the group mask with the new mask */
612 if (dropped || do_group)
613 fsnotify_recalc_group_mask(group);
619 /* match the get from fsnotify_find_mark_entry() */
620 fsnotify_put_mark(entry);
625 static int inotify_new_watch(struct fsnotify_group *group,
629 struct inotify_inode_mark_entry *tmp_ientry;
632 struct idr *idr = &group->inotify_data.idr;
633 spinlock_t *idr_lock = &group->inotify_data.idr_lock;
635 /* don't allow invalid bits: we don't want flags set */
636 mask = inotify_arg_to_mask(arg);
640 tmp_ientry = kmem_cache_alloc(inotify_inode_mark_cachep, GFP_KERNEL);
641 if (unlikely(!tmp_ientry))
644 fsnotify_init_mark(&tmp_ientry->fsn_entry, inotify_free_mark);
645 tmp_ientry->fsn_entry.mask = mask;
649 if (atomic_read(&group->inotify_data.user->inotify_watches) >= inotify_max_user_watches)
652 ret = inotify_add_to_idr(idr, idr_lock, &group->inotify_data.last_wd,
657 /* we are on the idr, now get on the inode */
658 ret = fsnotify_add_mark(&tmp_ientry->fsn_entry, group, inode, 0);
660 /* we failed to get on the inode, get off the idr */
661 inotify_remove_from_idr(group, tmp_ientry);
665 /* increment the number of watches the user has */
666 atomic_inc(&group->inotify_data.user->inotify_watches);
668 /* return the watch descriptor for this new entry */
669 ret = tmp_ientry->wd;
671 /* if this mark added a new event update the group mask */
672 if (mask & ~group->mask)
673 fsnotify_recalc_group_mask(group);
676 /* match the ref from fsnotify_init_markentry() */
677 fsnotify_put_mark(&tmp_ientry->fsn_entry);
682 static int inotify_update_watch(struct fsnotify_group *group, struct inode *inode, u32 arg)
687 /* try to update and existing watch with the new arg */
688 ret = inotify_update_existing_watch(group, inode, arg);
689 /* no mark present, try to add a new one */
691 ret = inotify_new_watch(group, inode, arg);
693 * inotify_new_watch could race with another thread which did an
694 * inotify_new_watch between the update_existing and the add watch
695 * here, go back and try to update an existing mark again.
703 static struct fsnotify_group *inotify_new_group(struct user_struct *user, unsigned int max_events)
705 struct fsnotify_group *group;
706 unsigned int grp_num;
708 /* fsnotify_obtain_group took a reference to group, we put this when we kill the file in the end */
709 grp_num = (INOTIFY_GROUP_NUM - atomic_inc_return(&inotify_grp_num));
710 group = fsnotify_obtain_group(grp_num, 0, &inotify_fsnotify_ops);
714 group->max_events = max_events;
716 spin_lock_init(&group->inotify_data.idr_lock);
717 idr_init(&group->inotify_data.idr);
718 group->inotify_data.last_wd = 0;
719 group->inotify_data.user = user;
720 group->inotify_data.fa = NULL;
726 /* inotify syscalls */
727 SYSCALL_DEFINE1(inotify_init1, int, flags)
729 struct fsnotify_group *group;
730 struct user_struct *user;
733 /* Check the IN_* constants for consistency. */
734 BUILD_BUG_ON(IN_CLOEXEC != O_CLOEXEC);
735 BUILD_BUG_ON(IN_NONBLOCK != O_NONBLOCK);
737 if (flags & ~(IN_CLOEXEC | IN_NONBLOCK))
740 user = get_current_user();
741 if (unlikely(atomic_read(&user->inotify_devs) >=
742 inotify_max_user_instances)) {
747 /* fsnotify_obtain_group took a reference to group, we put this when we kill the file in the end */
748 group = inotify_new_group(user, inotify_max_queued_events);
750 ret = PTR_ERR(group);
754 atomic_inc(&user->inotify_devs);
756 ret = anon_inode_getfd("inotify", &inotify_fops, group,
761 atomic_dec(&user->inotify_devs);
767 SYSCALL_DEFINE0(inotify_init)
769 return sys_inotify_init1(0);
772 SYSCALL_DEFINE3(inotify_add_watch, int, fd, const char __user *, pathname,
775 struct fsnotify_group *group;
779 int ret, fput_needed;
782 filp = fget_light(fd, &fput_needed);
786 /* verify that this is indeed an inotify instance */
787 if (unlikely(filp->f_op != &inotify_fops)) {
792 if (!(mask & IN_DONT_FOLLOW))
793 flags |= LOOKUP_FOLLOW;
794 if (mask & IN_ONLYDIR)
795 flags |= LOOKUP_DIRECTORY;
797 ret = inotify_find_inode(pathname, &path, flags);
801 /* inode held in place by reference to path; group by fget on fd */
802 inode = path.dentry->d_inode;
803 group = filp->private_data;
805 /* create/update an inode mark */
806 ret = inotify_update_watch(group, inode, mask);
809 fput_light(filp, fput_needed);
813 SYSCALL_DEFINE2(inotify_rm_watch, int, fd, __s32, wd)
815 struct fsnotify_group *group;
816 struct inotify_inode_mark_entry *ientry;
818 int ret = 0, fput_needed;
820 filp = fget_light(fd, &fput_needed);
824 /* verify that this is indeed an inotify instance */
826 if (unlikely(filp->f_op != &inotify_fops))
829 group = filp->private_data;
832 ientry = inotify_idr_find(group, wd);
833 if (unlikely(!ientry))
838 fsnotify_destroy_mark_by_entry(&ientry->fsn_entry);
840 /* match ref taken by inotify_idr_find */
841 fsnotify_put_mark(&ientry->fsn_entry);
844 fput_light(filp, fput_needed);
849 * inotify_user_setup - Our initialization function. Note that we cannnot return
850 * error because we have compiled-in VFS hooks. So an (unlikely) failure here
851 * must result in panic().
853 static int __init inotify_user_setup(void)
855 inotify_inode_mark_cachep = KMEM_CACHE(inotify_inode_mark_entry, SLAB_PANIC);
856 event_priv_cachep = KMEM_CACHE(inotify_event_private_data, SLAB_PANIC);
858 inotify_max_queued_events = 16384;
859 inotify_max_user_instances = 128;
860 inotify_max_user_watches = 8192;
864 module_init(inotify_user_setup);