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>
7Document updated 4 Jan 2015 by Zhang Zhen <zhenzhang.zhang@huawei.com>
8	--Deleted obsoleted interface, just refer to manpages for user interface.
9
10(i) Rationale
11
12Q: What is the design decision behind not tying the watch to the open fd of
13   the watched object?
14
15A: Watches are associated with an open inotify device, not an open file.
16   This solves the primary problem with dnotify: keeping the file open pins
17   the file and thus, worse, pins the mount.  Dnotify is therefore infeasible
18   for use on a desktop system with removable media as the media cannot be
19   unmounted.  Watching a file should not require that it be open.
20
21Q: What is the design decision behind using an-fd-per-instance as opposed to
22   an fd-per-watch?
23
24A: An fd-per-watch quickly consumes more file descriptors than are allowed,
25   more fd's than are feasible to manage, and more fd's than are optimally
26   select()-able.  Yes, root can bump the per-process fd limit and yes, users
27   can use epoll, but requiring both is a silly and extraneous requirement.
28   A watch consumes less memory than an open file, separating the number
29   spaces is thus sensible.  The current design is what user-space developers
30   want: Users initialize inotify, once, and add n watches, requiring but one
31   fd and no twiddling with fd limits.  Initializing an inotify instance two
32   thousand times is silly.  If we can implement user-space's preferences 
33   cleanly--and we can, the idr layer makes stuff like this trivial--then we 
34   should.
35
36   There are other good arguments.  With a single fd, there is a single
37   item to block on, which is mapped to a single queue of events.  The single
38   fd returns all watch events and also any potential out-of-band data.  If
39   every fd was a separate watch,
40
41   - There would be no way to get event ordering.  Events on file foo and
42     file bar would pop poll() on both fd's, but there would be no way to tell
43     which happened first.  A single queue trivially gives you ordering.  Such
44     ordering is crucial to existing applications such as Beagle.  Imagine
45     "mv a b ; mv b a" events without ordering.
46
47   - We'd have to maintain n fd's and n internal queues with state,
48     versus just one.  It is a lot messier in the kernel.  A single, linear
49     queue is the data structure that makes sense.
50
51   - User-space developers prefer the current API.  The Beagle guys, for
52     example, love it.  Trust me, I asked.  It is not a surprise: Who'd want
53     to manage and block on 1000 fd's via select?
54
55   - No way to get out of band data.
56
57   - 1024 is still too low.  ;-)
58
59   When you talk about designing a file change notification system that
60   scales to 1000s of directories, juggling 1000s of fd's just does not seem
61   the right interface.  It is too heavy.
62
63   Additionally, it _is_ possible to  more than one instance  and
64   juggle more than one queue and thus more than one associated fd.  There
65   need not be a one-fd-per-process mapping; it is one-fd-per-queue and a
66   process can easily want more than one queue.
67
68Q: Why the system call approach?
69
70A: The poor user-space interface is the second biggest problem with dnotify.
71   Signals are a terrible, terrible interface for file notification.  Or for
72   anything, for that matter.  The ideal solution, from all perspectives, is a
73   file descriptor-based one that allows basic file I/O and poll/select.
74   Obtaining the fd and managing the watches could have been done either via a
75   device file or a family of new system calls.  We decided to implement a
76   family of system calls because that is the preferred approach for new kernel
77   interfaces.  The only real difference was whether we wanted to use open(2)
78   and ioctl(2) or a couple of new system calls.  System calls beat ioctls.
79
80