1				   inotify
2	    a powerful yet simple file change notification system
3
4
5
6Document started 15 Mar 2005 by Robert Love <rml@novell.com>
7
8
9(i) User Interface
10
11Inotify is controlled by a set of three system calls and normal file I/O on a
12returned file descriptor.
13
14First step in using inotify is to initialise an inotify instance:
15
16	int fd = inotify_init ();
17
18Each instance is associated with a unique, ordered queue.
19
20Change events are managed by "watches".  A watch is an (object,mask) pair where
21the object is a file or directory and the mask is a bit mask of one or more
22inotify events that the application wishes to receive.  See <linux/inotify.h>
23for valid events.  A watch is referenced by a watch descriptor, or wd.
24
25Watches are added via a path to the file.
26
27Watches on a directory will return events on any files inside of the directory.
28
29Adding a watch is simple:
30
31	int wd = inotify_add_watch (fd, path, mask);
32
33Where "fd" is the return value from inotify_init(), path is the path to the
34object to watch, and mask is the watch mask (see <linux/inotify.h>).
35
36You can update an existing watch in the same manner, by passing in a new mask.
37
38An existing watch is removed via
39
40	int ret = inotify_rm_watch (fd, wd);
41
42Events are provided in the form of an inotify_event structure that is read(2)
43from a given inotify instance.  The filename is of dynamic length and follows
44the struct. It is of size len.  The filename is padded with null bytes to
45ensure proper alignment.  This padding is reflected in len.
46
47You can slurp multiple events by passing a large buffer, for example
48
49	size_t len = read (fd, buf, BUF_LEN);
50
51Where "buf" is a pointer to an array of "inotify_event" structures at least
52BUF_LEN bytes in size.  The above example will return as many events as are
53available and fit in BUF_LEN.
54
55Each inotify instance fd is also select()- and poll()-able.
56
57You can find the size of the current event queue via the standard FIONREAD
58ioctl on the fd returned by inotify_init().
59
60All watches are destroyed and cleaned up on close.
61
62
63(ii)
64
65Prototypes:
66
67	int inotify_init (void);
68	int inotify_add_watch (int fd, const char *path, __u32 mask);
69	int inotify_rm_watch (int fd, __u32 mask);
70
71
72(iii) Kernel Interface
73
74Inotify's kernel API consists a set of functions for managing watches and an
75event callback.
76
77To use the kernel API, you must first initialize an inotify instance with a set
78of inotify_operations.  You are given an opaque inotify_handle, which you use
79for any further calls to inotify.
80
81    struct inotify_handle *ih = inotify_init(my_event_handler);
82
83You must provide a function for processing events and a function for destroying
84the inotify watch.
85
86    void handle_event(struct inotify_watch *watch, u32 wd, u32 mask,
87    	              u32 cookie, const char *name, struct inode *inode)
88
89	watch - the pointer to the inotify_watch that triggered this call
90	wd - the watch descriptor
91	mask - describes the event that occurred
92	cookie - an identifier for synchronizing events
93	name - the dentry name for affected files in a directory-based event
94	inode - the affected inode in a directory-based event
95
96    void destroy_watch(struct inotify_watch *watch)
97
98You may add watches by providing a pre-allocated and initialized inotify_watch
99structure and specifying the inode to watch along with an inotify event mask.
100You must pin the inode during the call.  You will likely wish to embed the
101inotify_watch structure in a structure of your own which contains other
102information about the watch.  Once you add an inotify watch, it is immediately
103subject to removal depending on filesystem events.  You must grab a reference if
104you depend on the watch hanging around after the call.
105
106    inotify_init_watch(&my_watch->iwatch);
107    inotify_get_watch(&my_watch->iwatch);	// optional
108    s32 wd = inotify_add_watch(ih, &my_watch->iwatch, inode, mask);
109    inotify_put_watch(&my_watch->iwatch);	// optional
110
111You may use the watch descriptor (wd) or the address of the inotify_watch for
112other inotify operations.  You must not directly read or manipulate data in the
113inotify_watch.  Additionally, you must not call inotify_add_watch() more than
114once for a given inotify_watch structure, unless you have first called either
115inotify_rm_watch() or inotify_rm_wd().
116
117To determine if you have already registered a watch for a given inode, you may
118call inotify_find_watch(), which gives you both the wd and the watch pointer for
119the inotify_watch, or an error if the watch does not exist.
120
121    wd = inotify_find_watch(ih, inode, &watchp);
122
123You may use container_of() on the watch pointer to access your own data
124associated with a given watch.  When an existing watch is found,
125inotify_find_watch() bumps the refcount before releasing its locks.  You must
126put that reference with:
127
128    put_inotify_watch(watchp);
129
130Call inotify_find_update_watch() to update the event mask for an existing watch.
131inotify_find_update_watch() returns the wd of the updated watch, or an error if
132the watch does not exist.
133
134    wd = inotify_find_update_watch(ih, inode, mask);
135
136An existing watch may be removed by calling either inotify_rm_watch() or
137inotify_rm_wd().
138
139    int ret = inotify_rm_watch(ih, &my_watch->iwatch);
140    int ret = inotify_rm_wd(ih, wd);
141
142A watch may be removed while executing your event handler with the following:
143
144    inotify_remove_watch_locked(ih, iwatch);
145
146Call inotify_destroy() to remove all watches from your inotify instance and
147release it.  If there are no outstanding references, inotify_destroy() will call
148your destroy_watch op for each watch.
149
150    inotify_destroy(ih);
151
152When inotify removes a watch, it sends an IN_IGNORED event to your callback.
153You may use this event as an indication to free the watch memory.  Note that
154inotify may remove a watch due to filesystem events, as well as by your request.
155If you use IN_ONESHOT, inotify will remove the watch after the first event, at
156which point you may call the final inotify_put_watch.
157
158(iv) Kernel Interface Prototypes
159
160	struct inotify_handle *inotify_init(struct inotify_operations *ops);
161
162	inotify_init_watch(struct inotify_watch *watch);
163
164	s32 inotify_add_watch(struct inotify_handle *ih,
165		              struct inotify_watch *watch,
166			      struct inode *inode, u32 mask);
167
168	s32 inotify_find_watch(struct inotify_handle *ih, struct inode *inode,
169			       struct inotify_watch **watchp);
170
171	s32 inotify_find_update_watch(struct inotify_handle *ih,
172				      struct inode *inode, u32 mask);
173
174	int inotify_rm_wd(struct inotify_handle *ih, u32 wd);
175
176	int inotify_rm_watch(struct inotify_handle *ih,
177			     struct inotify_watch *watch);
178
179	void inotify_remove_watch_locked(struct inotify_handle *ih,
180					 struct inotify_watch *watch);
181
182	void inotify_destroy(struct inotify_handle *ih);
183
184	void get_inotify_watch(struct inotify_watch *watch);
185	void put_inotify_watch(struct inotify_watch *watch);
186
187
188(v) Internal Kernel Implementation
189
190Each inotify instance is represented by an inotify_handle structure.
191Inotify's userspace consumers also have an inotify_device which is
192associated with the inotify_handle, and on which events are queued.
193
194Each watch is associated with an inotify_watch structure.  Watches are chained
195off of each associated inotify_handle and each associated inode.
196
197See fs/inotify.c and fs/inotify_user.c for the locking and lifetime rules.
198
199
200(vi) Rationale
201
202Q: What is the design decision behind not tying the watch to the open fd of
203   the watched object?
204
205A: Watches are associated with an open inotify device, not an open file.
206   This solves the primary problem with dnotify: keeping the file open pins
207   the file and thus, worse, pins the mount.  Dnotify is therefore infeasible
208   for use on a desktop system with removable media as the media cannot be
209   unmounted.  Watching a file should not require that it be open.
210
211Q: What is the design decision behind using an-fd-per-instance as opposed to
212   an fd-per-watch?
213
214A: An fd-per-watch quickly consumes more file descriptors than are allowed,
215   more fd's than are feasible to manage, and more fd's than are optimally
216   select()-able.  Yes, root can bump the per-process fd limit and yes, users
217   can use epoll, but requiring both is a silly and extraneous requirement.
218   A watch consumes less memory than an open file, separating the number
219   spaces is thus sensible.  The current design is what user-space developers
220   want: Users initialize inotify, once, and add n watches, requiring but one
221   fd and no twiddling with fd limits.  Initializing an inotify instance two
222   thousand times is silly.  If we can implement user-space's preferences
223   cleanly--and we can, the idr layer makes stuff like this trivial--then we
224   should.
225
226   There are other good arguments.  With a single fd, there is a single
227   item to block on, which is mapped to a single queue of events.  The single
228   fd returns all watch events and also any potential out-of-band data.  If
229   every fd was a separate watch,
230
231   - There would be no way to get event ordering.  Events on file foo and
232     file bar would pop poll() on both fd's, but there would be no way to tell
233     which happened first.  A single queue trivially gives you ordering.  Such
234     ordering is crucial to existing applications such as Beagle.  Imagine
235     "mv a b ; mv b a" events without ordering.
236
237   - We'd have to maintain n fd's and n internal queues with state,
238     versus just one.  It is a lot messier in the kernel.  A single, linear
239     queue is the data structure that makes sense.
240
241   - User-space developers prefer the current API.  The Beagle guys, for
242     example, love it.  Trust me, I asked.  It is not a surprise: Who'd want
243     to manage and block on 1000 fd's via select?
244
245   - No way to get out of band data.
246
247   - 1024 is still too low.  ;-)
248
249   When you talk about designing a file change notification system that
250   scales to 1000s of directories, juggling 1000s of fd's just does not seem
251   the right interface.  It is too heavy.
252
253   Additionally, it _is_ possible to  more than one instance  and
254   juggle more than one queue and thus more than one associated fd.  There
255   need not be a one-fd-per-process mapping; it is one-fd-per-queue and a
256   process can easily want more than one queue.
257
258Q: Why the system call approach?
259
260A: The poor user-space interface is the second biggest problem with dnotify.
261   Signals are a terrible, terrible interface for file notification.  Or for
262   anything, for that matter.  The ideal solution, from all perspectives, is a
263   file descriptor-based one that allows basic file I/O and poll/select.
264   Obtaining the fd and managing the watches could have been done either via a
265   device file or a family of new system calls.  We decided to implement a
266   family of system calls because that is the preferred approach for new kernel
267   interfaces.  The only real difference was whether we wanted to use open(2)
268   and ioctl(2) or a couple of new system calls.  System calls beat ioctls.
269
270