				    inotify
	     a powerful yet simple file change notification system


Document started 15 Mar 2005 by Robert Love <rml@novell.com>

(i) User Interface

Inotify is controlled by a set of three sys calls 

First step in using inotify is to initialise an inotify instance

	int fd = inotify_init ();

Change events are managed by "watches".  A watch is an (object,mask) pair where
the object is a file or directory and the mask is a bit mask of one or more
inotify events that the application wishes to receive.  See <linux/inotify.h>
for valid events.  A watch is referenced by a watch descriptor, or wd.

Watches are added via a path to the file.

Watches on a directory will return events on any files inside of the directory.

Adding a watch is simple,

	int wd = inotify_add_watch (fd, path, mask);

You can add a large number of files via something like

	for each file to watch {
		int wd = inotify_add_watch (fd, file, mask);
	}

You can update an existing watch in the same manner, by passing in a new mask.

An existing watch is removed via the INOTIFY_IGNORE ioctl, for example

	inotify_rm_watch (fd, wd);

Events are provided in the form of an inotify_event structure that is read(2)
from a inotify instance fd.  The filename is of dynamic length and follows the 
struct. It is of size len.  The filename is padded with null bytes to ensure 
proper alignment.  This padding is reflected in len.

You can slurp multiple events by passing a large buffer, for example

	size_t len = read (fd, buf, BUF_LEN);

Will return as many events as are available and fit in BUF_LEN.

each inotify instance fd is also select()- and poll()-able.

You can find the size of the current event queue via the FIONREAD ioctl.

All watches are destroyed and cleaned up on close.


(ii) Internal Kernel Implementation

Each open inotify instance is associated with an inotify_device structure.

Each watch is associated with an inotify_watch structure.  Watches are chained
off of each associated device and each associated inode.

See fs/inotify.c for the locking and lifetime rules.


(iii) Rationale

Q: What is the design decision behind not tying the watch to the open fd of
   the watched object?

A: Watches are associated with an open inotify device, not an open file.
   This solves the primary problem with dnotify: keeping the file open pins
   the file and thus, worse, pins the mount.  Dnotify is therefore infeasible
   for use on a desktop system with removable media as the media cannot be
   unmounted.

Q: What is the design decision behind using an-fd-per-device as opposed to
   an fd-per-watch?

A: An fd-per-watch quickly consumes more file descriptors than are allowed,
   more fd's than are feasible to manage, and more fd's than are optimally
   select()-able.  Yes, root can bump the per-process fd limit and yes, users
   can use epoll, but requiring both is a silly and extraneous requirement.
   A watch consumes less memory than an open file, separating the number
   spaces is thus sensible.  The current design is what user-space developers
   want: Users initialize inotify, once, and add n watches, requiring but one fd
   and no twiddling with fd limits.  Initializing an inotify instance two
   thousand times is silly.  If we can implement user-space's preferences 
   cleanly--and we can, the idr layer makes stuff like this trivial--then we 
   should.

   There are other good arguments.  With a single fd, there is a single
   item to block on, which is mapped to a single queue of events.  The single
   fd returns all watch events and also any potential out-of-band data.  If
   every fd was a separate watch,

   - There would be no way to get event ordering.  Events on file foo and
     file bar would pop poll() on both fd's, but there would be no way to tell
     which happened first.  A single queue trivially gives you ordering.  Such
     ordering is crucial to existing applications such as Beagle.  Imagine
     "mv a b ; mv b a" events without ordering.

   - We'd have to maintain n fd's and n internal queues with state,
     versus just one.  It is a lot messier in the kernel.  A single, linear
     queue is the data structure that makes sense.

   - User-space developers prefer the current API.  The Beagle guys, for
     example, love it.  Trust me, I asked.  It is not a surprise: Who'd want
     to manage and block on 1000 fd's via select?

   - You'd have to manage the fd's, as an example: Call close() when you
     received a delete event.

   - No way to get out of band data.

   - 1024 is still too low.  ;-)

   When you talk about designing a file change notification system that
   scales to 1000s of directories, juggling 1000s of fd's just does not seem
   the right interface.  It is too heavy.

Q: Why the system call approach?

A: The poor user-space interface is the second biggest problem with dnotify.
   Signals are a terrible, terrible interface for file notification.  Or for
   anything, for that matter.  The ideal solution, from all perspectives, is a
   file descriptor-based one that allows basic file I/O and poll/select.
   Obtaining the fd and managing the watches could have been done either via a
   device file or a family of new system calls.  We decided to implement a
   family of system calls because that is the preffered approach for new kernel
   features and it means our user interface requirements.

   Additionally, it _is_ possible to  more than one instance  and
   juggle more than one queue and thus more than one associated fd.