2 .\" Copyright (C) 2006, 2014 Michael Kerrisk <mtk.manpages@gmail.com>
3 .\" Copyright (C) 2014 Heinrich Schuchardt <xypron.glpk@gmx.de>
5 .\" %%%LICENSE_START(VERBATIM)
6 .\" Permission is granted to make and distribute verbatim copies of this
7 .\" manual provided the copyright notice and this permission notice are
8 .\" preserved on all copies.
10 .\" Permission is granted to copy and distribute modified versions of this
11 .\" manual under the conditions for verbatim copying, provided that the
12 .\" entire resulting derived work is distributed under the terms of a
13 .\" permission notice identical to this one.
15 .\" Since the Linux kernel and libraries are constantly changing, this
16 .\" manual page may be incorrect or out-of-date. The author(s) assume no
17 .\" responsibility for errors or omissions, or for damages resulting from
18 .\" the use of the information contained herein. The author(s) may not
19 .\" have taken the same level of care in the production of this manual,
20 .\" which is licensed free of charge, as they might when working
23 .\" Formatted or processed versions of this manual, if unaccompanied by
24 .\" the source, must acknowledge the copyright and authors of this work.
27 .TH INOTIFY 7 2014-07-08 "Linux" "Linux Programmer's Manual"
29 inotify \- monitoring filesystem events
33 API provides a mechanism for monitoring filesystem events.
34 Inotify can be used to monitor individual files,
35 or to monitor directories.
36 When a directory is monitored, inotify will return events
37 for the directory itself, and for files inside the directory.
39 The following system calls are used with this API:
42 creates an inotify instance and returns a file descriptor
43 referring to the inotify instance.
50 argument that provides access to some extra functionality.
52 .BR inotify_add_watch (2)
53 manipulates the "watch list" associated with an inotify instance.
54 Each item ("watch") in the watch list specifies the pathname of
56 along with some set of events that the kernel should monitor for the
57 file referred to by that pathname.
58 .BR inotify_add_watch (2)
59 either creates a new watch item, or modifies an existing watch.
60 Each watch has a unique "watch descriptor", an integer
62 .BR inotify_add_watch (2)
63 when the watch is created.
65 When events occur for monitored files and directories,
66 those events are made available to the application as structured data that
67 can be read from the inotify file descriptor using
71 .BR inotify_rm_watch (2)
72 removes an item from an inotify watch list.
74 When all file descriptors referring to an inotify
75 instance have been closed (using
77 the underlying object and its resources are
78 freed for reuse by the kernel;
79 all associated watches are automatically freed.
81 With careful programming,
82 an application can use inotify to efficiently monitor and cache
83 the state of a set of filesystem objects.
84 However, robust applications should allow for the fact that bugs
85 in the monitoring logic or races of the kind described below
86 may leave the cache inconsistent with the filesystem state.
87 It is probably wise to to do some consistency checking,
88 and rebuild the cache when inconsistencies are detected.
89 .SS Reading events from an inotify file descriptor
90 To determine what events have occurred, an application
92 from the inotify file descriptor.
93 If no events have so far occurred, then,
94 assuming a blocking file descriptor,
96 will block until at least one event occurs
97 (unless interrupted by a signal,
98 in which case the call fails with the error
105 returns a buffer containing one or more of the following structures:
109 struct inotify_event {
110 int wd; /* Watch descriptor */
111 .\" FIXME . The type of the 'wd' field should probably be "int32_t".
112 .\" I submitted a patch to fix this. See the LKML thread
113 .\" "[patch] Fix type errors in inotify interfaces", 18 Nov 2008
114 .\" Glibc bug filed: http://sources.redhat.com/bugzilla/show_bug.cgi?id=7040
115 uint32_t mask; /* Mask describing event */
116 uint32_t cookie; /* Unique cookie associating related
117 events (for rename(2)) */
118 uint32_t len; /* Size of \fIname\fP field */
119 char name[]; /* Optional null-terminated name */
125 identifies the watch for which this event occurs.
126 It is one of the watch descriptors returned by a previous call to
127 .BR inotify_add_watch (2).
130 contains bits that describe the event that occurred (see below).
133 is a unique integer that connects related events.
134 Currently this is used only for rename events, and
135 allows the resulting pair of
139 events to be connected by the application.
140 For all other event types,
146 field is present only when an event is returned
147 for a file inside a watched directory;
148 it identifies the file pathname relative to the watched directory.
149 This pathname is null-terminated,
150 and may include further null bytes (\(aq\\0\(aq) to align subsequent reads to a
151 suitable address boundary.
155 field counts all of the bytes in
157 including the null bytes;
161 .IR "sizeof(struct inotify_event)+len" .
163 The behavior when the buffer given to
165 is too small to return information about the next event depends
166 on the kernel version: in kernels before 2.6.21,
168 returns 0; since kernel 2.6.21,
172 Specifying a buffer of size
174 sizeof(struct inotify_event) + NAME_MAX + 1
176 will be sufficient to read at least one event.
179 .BR inotify_add_watch (2)
185 structure returned when
187 an inotify file descriptor are both bit masks identifying
189 The following bits can be specified in
192 .BR inotify_add_watch (2)
193 and may be returned in the
200 File was accessed (e.g.,
205 Metadata changed\(emfor example, permissions (e.g.,
211 link count (since Linux 2.6.25; e.g.,
216 and user/group ID (e.g.,
219 .BR IN_CLOSE_WRITE " (*)"
220 File opened for writing was closed.
222 .BR IN_CLOSE_NOWRITE " (*)"
223 File not opened for writing was closed.
226 File/directory created in watched directory (e.g.,
233 on a UNIX domain socket).
236 File/directory deleted from watched directory.
239 Watched file/directory was itself deleted.
240 (This event also occurs if an object is moved to another filesystem,
243 in effect copies the file to the other filesystem and
244 then deletes it from the original filesystem.)
247 event will subsequently be generated for the watch descriptor.
250 File was modified (e.g.,
255 Watched file/directory was itself moved.
257 .BR IN_MOVED_FROM " (+)"
258 Generated for the directory containing the old filename
259 when a file is renamed.
261 .BR IN_MOVED_TO " (+)"
262 Generated for the directory containing the new filename
263 when a file is renamed.
269 When monitoring a directory:
271 the events marked above with an asterisk (*) can occur both
272 for the directory itself and for objects inside the directory; and
274 the events marked with a plus sign (+) occur only for objects
275 inside the directory (not for the directory itself).
277 When events are generated for objects inside a watched directory, the
279 field in the returned
281 structure identifies the name of the file within the directory.
285 macro is defined as a bit mask of all of the above events.
286 This macro can be used as the
288 argument when calling
289 .BR inotify_add_watch (2).
291 Two additional convenience macros are defined:
296 .BR "IN_MOVED_FROM | IN_MOVED_TO" .
300 .BR "IN_CLOSE_WRITE | IN_CLOSE_NOWRITE" .
303 The following further bits can be specified in
306 .BR inotify_add_watch (2):
309 .BR IN_DONT_FOLLOW " (since Linux 2.6.15)"
312 if it is a symbolic link.
314 .BR IN_EXCL_UNLINK " (since Linux 2.6.36)"
315 .\" commit 8c1934c8d70b22ca8333b216aec6c7d09fdbd6a6
316 By default, when watching events on the children of a directory,
317 events are generated for children even after they have been unlinked
319 This can result in large numbers of uninteresting events for
320 some applications (e.g., if watching
322 in which many applications create temporary files whose
323 names are immediately unlinked).
326 changes the default behavior,
327 so that events are not generated for children after
328 they have been unlinked from the watched directory.
331 If a watch instance already exists for the filesystem object corresponding to
333 add (OR) the events in
335 to the watch mask (instead of replacing the mask).
338 Monitor the filesystem object corresponding to
340 for one event, then remove from
343 .BR IN_ONLYDIR " (since Linux 2.6.15)"
346 if it is a directory.
347 Using this flag provides an application with a race-free way of
348 ensuring that the monitored objeect is a directory.
351 The following bits may be set in the
358 Watch was removed explicitly
359 .RB ( inotify_rm_watch (2))
360 or automatically (file was deleted, or filesystem was unmounted).
364 Subject of this event is a directory.
367 Event queue overflowed
369 is \-1 for this event).
372 Filesystem containing watched object was unmounted.
375 event will subsequently be generated for the watch descriptor.
378 Suppose an application is watching the directory
383 The examples below show some events that will be generated
384 for these two objects.
387 fd = open("dir/myfile", O_RDWR);
395 read(fd, buf, count);
403 write(fd, buf, count);
428 Suppose an application is watching the directories
434 The following examples show some events that may be generated.
437 link("dir1/myfile", "dir2/new");
447 rename("dir1/myfile", "dir2/myfile");
464 events will have the same
473 are (the only) links to the same file, and an application is watching
479 Executing the following calls in the order given below will generate
480 the following events:
488 (because its link count changes)
508 Suppose an application is watching the directory
510 and (the empty) directory
512 The following examples show some events that may be generated.
515 mkdir("dir/new", mode);
517 .B "IN_CREATE | IN_ISDIR"
529 .B "IN_DELETE | IN_ISDIR"
534 The following interfaces can be used to limit the amount of
535 kernel memory consumed by inotify:
537 .I /proc/sys/fs/inotify/max_queued_events
538 The value in this file is used when an application calls
540 to set an upper limit on the number of events that can be
541 queued to the corresponding inotify instance.
542 Events in excess of this limit are dropped, but an
544 event is always generated.
546 .I /proc/sys/fs/inotify/max_user_instances
547 This specifies an upper limit on the number of inotify instances
548 that can be created per real user ID.
550 .I /proc/sys/fs/inotify/max_user_watches
551 This specifies an upper limit on the number of watches
552 that can be created per real user ID.
554 Inotify was merged into the 2.6.13 Linux kernel.
555 The required library interfaces were added to glibc in version 2.4.
556 .RB ( IN_DONT_FOLLOW ,
560 were added in glibc version 2.5.)
562 The inotify API is Linux-specific.
564 Inotify file descriptors can be monitored using
569 When an event is available, the file descriptor indicates as readable.
572 signal-driven I/O notification is available for inotify file descriptors;
573 see the discussion of
585 structure (described in
587 that is passed to the signal handler has the following fields set:
589 is set to the inotify file descriptor number;
591 is set to the signal number;
600 If successive output inotify events produced on the
601 inotify file descriptor are identical (same
607 then they are coalesced into a single event if the
608 older event has not yet been read (but see BUGS).
609 This reduces the amount of kernel memory required for the event queue,
610 but also means that an application can't use inotify to reliably count
613 The events returned by reading from an inotify file descriptor
614 form an ordered queue.
615 Thus, for example, it is guaranteed that when renaming from
616 one directory to another, events will be produced in the
617 correct order on the inotify file descriptor.
622 returns the number of bytes available to read from an
623 inotify file descriptor.
624 .SS Limitations and caveats
625 The inotify API provides no information about the user or process that
626 triggered the inotify event.
627 In particular, there is no easy
628 way for a process that is monitoring events via inotify
629 to distinguish events that it triggers
630 itself from those that are triggered by other processes.
632 Inotify reports only events that a user-space program triggers through
634 As a result, it does not catch remote events that occur
635 on network filesystems.
636 (Applications must fall back to polling the filesystem
637 to catch such events.)
638 Furthermore, various pseudo-filesystems such as
643 are not monitorable with inotify.
645 The inotify API does not report file accesses and modifications that
652 The inotify API identifies affected files by filename.
653 However, by the time an application processes an inotify event,
654 the filename may already have been deleted or renamed.
656 The inotify API identifies events via watch descriptors.
657 It is the application's responsibility to cache a mapping
658 (if one is needed) between watch descriptors and pathnames.
659 Be aware that directory renamings may affect multiple cached pathnames.
661 Inotify monitoring of directories is not recursive:
662 to monitor subdirectories under a directory,
663 additional watches must be created.
664 This can take a significant amount time for large directory trees.
666 If monitoring an entire directory subtree,
667 and a new subdirectory is created in that tree or an existing directory
668 is renamed into that tree,
669 be aware that by the time you create a watch for the new subdirectory,
670 new files (and subdirectories) may already exist inside the subdirectory.
671 Therefore, you may want to scan the contents of the subdirectory
672 immediately after adding the watch (and, if desired,
673 recursively add watches for any subdirectories that it contains).
675 Note that the event queue can overflow.
676 In this case, events are lost.
677 Robust applications should handle the possibility of
678 lost events gracefully.
679 For example, it may be necessary to rebuild part or all of
680 the application cache.
681 (One simple, but possibly expensive,
682 approach is to close the inotify file descriptor, empty the cache,
683 create a new inotify file descriptor,
684 and then re-create watches and cache entries
685 for the objects to be monitored.)
686 .SS Dealing with rename() events
691 event pair that is generated by
693 can be matched up via their shared cookie value.
694 However, the task of matching has some challenges.
696 These two events are usually consecutive in the event stream available
697 when reading from the inotify file descriptor.
698 However, this is not guaranteed.
699 If multiple processes are triggering events for monitored objects,
700 then (on rare occasions) an arbitrary number of
701 other events may appear between the
706 Furthermore, it is not guaranteed that the event pair is atomically
707 inserted into the queue: there may be a brief interval where the
709 has appeared, but the
717 event pair generated by
719 is thus inherently racy.
720 (Don't forget that if an object is renamed outside of a monitored directory,
721 there may not even be an
724 Heuristic approaches (e.g., assume the events are always consecutive)
725 can be used to ensure a match in most cases,
726 but will inevitably miss some cases,
727 causing the application to perceive the
731 events as being unrelated.
732 If watch descriptors are destroyed and re-created as a result,
733 then those watch descriptors will be inconsistent with
734 the watch descriptors in any pending events.
735 (Re-creating the inotify file descriptor and rebuilding the cache may
736 be useful to deal with this scenario.)
738 Applications should also allow for the possibility that the
740 event was the last event that could fit in the buffer
741 returned by the current call to
745 event might be fetched only on the next
747 which should be done with a (small) timeout to allow for the fact that
749 .BR IN_MOVED_FROM - IN_MOVED_TO
750 event pair is not atomic,
751 and also the possibility that there may not be any
755 .\" FIXME . kernel commit 611da04f7a31b2208e838be55a42c7a1310ae321
756 .\" implies that unmount events were buggy 2.6.11 to 2.6.36
758 In kernels before 2.6.16, the
763 As originally designed and implemented, the
765 flag did not cause an
767 event to be generated when the watch was dropped after one event.
768 However, as an unintended effect of other changes,
769 since Linux 2.6.36, an
771 event is generated in this case.
773 Before kernel 2.6.25,
774 .\" commit 1c17d18e3775485bf1e0ce79575eb637a94494a2
775 the kernel code that was intended to coalesce successive identical events
776 (i.e., the two most recent events could potentially be coalesced
777 if the older had not yet been read)
778 instead checked if the most recent event could be coalesced with the
782 When a watch descriptor is removed by calling
783 .BR inotify_rm_watch (2)
784 (or because a watch file is deleted or the filesystem
785 that contains it is unmounted),
786 any pending unread events for that watch descriptor remain available to read.
787 As watch descriptors are subsequently allocated with
788 .BR inotify_add_watch (2),
789 the kernel cycles through the range of possible watch descriptors (0 to
792 When allocating a free watch descriptor, no check is made to see whether that
793 watch descriptor number has any pending unread events in the inotify queue.
794 Thus, it can happen that a watch descriptor is reallocated even
795 when pending unread events exist for a previous incarnation of
796 that watch descriptor number, with the result that the application
797 might then read those events and interpret them as belonging to
798 the file associated with the newly recycled watch descriptor.
799 In practice, the likelihood of hitting this bug may be extremely low,
800 since it requires that an application cycle through
803 release a watch descriptor while leaving unread events for that
804 watch descriptor in the queue in the queue,
805 and then recycle that watch descriptor.
806 For this reason, and because there have been no reports
807 of the bug occurring in real-world applications,
809 .\" FIXME . https://bugzilla.kernel.org/show_bug.cgi?id=77111
810 no kernel changes have yet been made to eliminate this possible bug.
812 The following program demonstrates the usage of the inotify API.
813 It marks the directories passed as a command-line arguments
814 and waits for events of type
820 The following output was recorded while editing the file
821 .I /home/user/temp/foo
822 and listing directory
824 Before the file and the directory were opened,
827 After the file was closed, an
830 After the directory was closed, an
833 Execution of the program ended when the user pressed the ENTER key.
837 $ \fB./a.out /tmp /home/user/temp\fP
838 Press enter key to terminate.
839 Listening for events.
840 IN_OPEN: /home/user/temp/foo [file]
841 IN_CLOSE_WRITE: /home/user/temp/foo [file]
842 IN_OPEN: /tmp/ [directory]
843 IN_CLOSE_NOWRITE: /tmp/ [directory]
845 Listening for events stopped.
854 #include <sys/inotify.h>
857 /* Read all available inotify events from the file descriptor 'fd'.
858 wd is the table of watch descriptors for the directories in argv.
859 argc is the length of wd and argv.
860 argv is the list of watched directories.
861 Entry 0 of wd and argv is unused. */
864 handle_events(int fd, int *wd, int argc, char* argv[])
866 /* Some systems cannot read integer variables if they are not
867 properly aligned. On other systems, incorrect alignment may
868 decrease performance. Hence, the buffer used for reading from
869 the inotify file descriptor should have the same alignment as
870 struct inotify_event. */
873 __attribute__ ((aligned(__alignof__(struct inotify_event))));
874 const struct inotify_event *event;
879 /* Loop while events can be read from inotify file descriptor. */
883 /* Read some events. */
885 len = read(fd, buf, sizeof buf);
886 if (len == \-1 && errno != EAGAIN) {
891 /* If the nonblocking read() found no events to read, then
892 it returns \-1 with errno set to EAGAIN. In that case,
898 /* Loop over all events in the buffer */
900 for (ptr = buf; ptr < buf + len;
901 ptr += sizeof(struct inotify_event) + event\->len) {
903 event = (const struct inotify_event *) ptr;
905 /* Print event type */
907 if (event\->mask & IN_OPEN)
909 if (event\->mask & IN_CLOSE_NOWRITE)
910 printf("IN_CLOSE_NOWRITE: ");
911 if (event\->mask & IN_CLOSE_WRITE)
912 printf("IN_CLOSE_WRITE: ");
914 /* Print the name of the watched directory */
916 for (i = 1; i < argc; ++i) {
917 if (wd[i] == event\->wd) {
918 printf("%s/", argv[i]);
923 /* Print the name of the file */
926 printf("%s", event\->name);
928 /* Print type of filesystem object */
930 if (event\->mask & IN_ISDIR)
931 printf(" [directory]\\n");
933 printf(" [file]\\n");
939 main(int argc, char* argv[])
945 struct pollfd fds[2];
948 printf("Usage: %s PATH [PATH ...]\\n", argv[0]);
952 printf("Press ENTER key to terminate.\\n");
954 /* Create the file descriptor for accessing the inotify API */
956 fd = inotify_init1(IN_NONBLOCK);
958 perror("inotify_init1");
962 /* Allocate memory for watch descriptors */
964 wd = calloc(argc, sizeof(int));
970 /* Mark directories for events
972 \- file was closed */
974 for (i = 1; i < argc; i++) {
975 wd[i] = inotify_add_watch(fd, argv[i],
978 fprintf(stderr, "Cannot watch '%s'\\n", argv[i]);
979 perror("inotify_add_watch");
984 /* Prepare for polling */
990 fds[0].fd = STDIN_FILENO;
991 fds[0].events = POLLIN;
996 fds[1].events = POLLIN;
998 /* Wait for events and/or terminal input */
1000 printf("Listening for events.\\n");
1002 poll_num = poll(fds, nfds, \-1);
1003 if (poll_num == \-1) {
1012 if (fds[0].revents & POLLIN) {
1014 /* Console input is available. Empty stdin and quit */
1016 while (read(STDIN_FILENO, &buf, 1) > 0 && buf != '\\n')
1021 if (fds[1].revents & POLLIN) {
1023 /* Inotify events are available */
1025 handle_events(fd, wd, argc, argv);
1030 printf("Listening for events stopped.\\n");
1032 /* Close inotify file descriptor */
1041 .BR inotifywait (1),
1042 .BR inotifywatch (1),
1043 .BR inotify_add_watch (2),
1044 .BR inotify_init (2),
1045 .BR inotify_init1 (2),
1046 .BR inotify_rm_watch (2),
1051 .IR Documentation/filesystems/inotify.txt
1052 in the Linux kernel source tree