inotify.txt 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269
  1. inotify
  2. a powerful yet simple file change notification system
  3. Document started 15 Mar 2005 by Robert Love <rml@novell.com>
  4. (i) User Interface
  5. Inotify is controlled by a set of three system calls and normal file I/O on a
  6. returned file descriptor.
  7. First step in using inotify is to initialise an inotify instance:
  8. int fd = inotify_init ();
  9. Each instance is associated with a unique, ordered queue.
  10. Change events are managed by "watches". A watch is an (object,mask) pair where
  11. the object is a file or directory and the mask is a bit mask of one or more
  12. inotify events that the application wishes to receive. See <linux/inotify.h>
  13. for valid events. A watch is referenced by a watch descriptor, or wd.
  14. Watches are added via a path to the file.
  15. Watches on a directory will return events on any files inside of the directory.
  16. Adding a watch is simple:
  17. int wd = inotify_add_watch (fd, path, mask);
  18. Where "fd" is the return value from inotify_init(), path is the path to the
  19. object to watch, and mask is the watch mask (see <linux/inotify.h>).
  20. You can update an existing watch in the same manner, by passing in a new mask.
  21. An existing watch is removed via
  22. int ret = inotify_rm_watch (fd, wd);
  23. Events are provided in the form of an inotify_event structure that is read(2)
  24. from a given inotify instance. The filename is of dynamic length and follows
  25. the struct. It is of size len. The filename is padded with null bytes to
  26. ensure proper alignment. This padding is reflected in len.
  27. You can slurp multiple events by passing a large buffer, for example
  28. size_t len = read (fd, buf, BUF_LEN);
  29. Where "buf" is a pointer to an array of "inotify_event" structures at least
  30. BUF_LEN bytes in size. The above example will return as many events as are
  31. available and fit in BUF_LEN.
  32. Each inotify instance fd is also select()- and poll()-able.
  33. You can find the size of the current event queue via the standard FIONREAD
  34. ioctl on the fd returned by inotify_init().
  35. All watches are destroyed and cleaned up on close.
  36. (ii)
  37. Prototypes:
  38. int inotify_init (void);
  39. int inotify_add_watch (int fd, const char *path, __u32 mask);
  40. int inotify_rm_watch (int fd, __u32 mask);
  41. (iii) Kernel Interface
  42. Inotify's kernel API consists a set of functions for managing watches and an
  43. event callback.
  44. To use the kernel API, you must first initialize an inotify instance with a set
  45. of inotify_operations. You are given an opaque inotify_handle, which you use
  46. for any further calls to inotify.
  47. struct inotify_handle *ih = inotify_init(my_event_handler);
  48. You must provide a function for processing events and a function for destroying
  49. the inotify watch.
  50. void handle_event(struct inotify_watch *watch, u32 wd, u32 mask,
  51. u32 cookie, const char *name, struct inode *inode)
  52. watch - the pointer to the inotify_watch that triggered this call
  53. wd - the watch descriptor
  54. mask - describes the event that occurred
  55. cookie - an identifier for synchronizing events
  56. name - the dentry name for affected files in a directory-based event
  57. inode - the affected inode in a directory-based event
  58. void destroy_watch(struct inotify_watch *watch)
  59. You may add watches by providing a pre-allocated and initialized inotify_watch
  60. structure and specifying the inode to watch along with an inotify event mask.
  61. You must pin the inode during the call. You will likely wish to embed the
  62. inotify_watch structure in a structure of your own which contains other
  63. information about the watch. Once you add an inotify watch, it is immediately
  64. subject to removal depending on filesystem events. You must grab a reference if
  65. you depend on the watch hanging around after the call.
  66. inotify_init_watch(&my_watch->iwatch);
  67. inotify_get_watch(&my_watch->iwatch); // optional
  68. s32 wd = inotify_add_watch(ih, &my_watch->iwatch, inode, mask);
  69. inotify_put_watch(&my_watch->iwatch); // optional
  70. You may use the watch descriptor (wd) or the address of the inotify_watch for
  71. other inotify operations. You must not directly read or manipulate data in the
  72. inotify_watch. Additionally, you must not call inotify_add_watch() more than
  73. once for a given inotify_watch structure, unless you have first called either
  74. inotify_rm_watch() or inotify_rm_wd().
  75. To determine if you have already registered a watch for a given inode, you may
  76. call inotify_find_watch(), which gives you both the wd and the watch pointer for
  77. the inotify_watch, or an error if the watch does not exist.
  78. wd = inotify_find_watch(ih, inode, &watchp);
  79. You may use container_of() on the watch pointer to access your own data
  80. associated with a given watch. When an existing watch is found,
  81. inotify_find_watch() bumps the refcount before releasing its locks. You must
  82. put that reference with:
  83. put_inotify_watch(watchp);
  84. Call inotify_find_update_watch() to update the event mask for an existing watch.
  85. inotify_find_update_watch() returns the wd of the updated watch, or an error if
  86. the watch does not exist.
  87. wd = inotify_find_update_watch(ih, inode, mask);
  88. An existing watch may be removed by calling either inotify_rm_watch() or
  89. inotify_rm_wd().
  90. int ret = inotify_rm_watch(ih, &my_watch->iwatch);
  91. int ret = inotify_rm_wd(ih, wd);
  92. A watch may be removed while executing your event handler with the following:
  93. inotify_remove_watch_locked(ih, iwatch);
  94. Call inotify_destroy() to remove all watches from your inotify instance and
  95. release it. If there are no outstanding references, inotify_destroy() will call
  96. your destroy_watch op for each watch.
  97. inotify_destroy(ih);
  98. When inotify removes a watch, it sends an IN_IGNORED event to your callback.
  99. You may use this event as an indication to free the watch memory. Note that
  100. inotify may remove a watch due to filesystem events, as well as by your request.
  101. If you use IN_ONESHOT, inotify will remove the watch after the first event, at
  102. which point you may call the final inotify_put_watch.
  103. (iv) Kernel Interface Prototypes
  104. struct inotify_handle *inotify_init(struct inotify_operations *ops);
  105. inotify_init_watch(struct inotify_watch *watch);
  106. s32 inotify_add_watch(struct inotify_handle *ih,
  107. struct inotify_watch *watch,
  108. struct inode *inode, u32 mask);
  109. s32 inotify_find_watch(struct inotify_handle *ih, struct inode *inode,
  110. struct inotify_watch **watchp);
  111. s32 inotify_find_update_watch(struct inotify_handle *ih,
  112. struct inode *inode, u32 mask);
  113. int inotify_rm_wd(struct inotify_handle *ih, u32 wd);
  114. int inotify_rm_watch(struct inotify_handle *ih,
  115. struct inotify_watch *watch);
  116. void inotify_remove_watch_locked(struct inotify_handle *ih,
  117. struct inotify_watch *watch);
  118. void inotify_destroy(struct inotify_handle *ih);
  119. void get_inotify_watch(struct inotify_watch *watch);
  120. void put_inotify_watch(struct inotify_watch *watch);
  121. (v) Internal Kernel Implementation
  122. Each inotify instance is represented by an inotify_handle structure.
  123. Inotify's userspace consumers also have an inotify_device which is
  124. associated with the inotify_handle, and on which events are queued.
  125. Each watch is associated with an inotify_watch structure. Watches are chained
  126. off of each associated inotify_handle and each associated inode.
  127. See fs/inotify.c and fs/inotify_user.c for the locking and lifetime rules.
  128. (vi) Rationale
  129. Q: What is the design decision behind not tying the watch to the open fd of
  130. the watched object?
  131. A: Watches are associated with an open inotify device, not an open file.
  132. This solves the primary problem with dnotify: keeping the file open pins
  133. the file and thus, worse, pins the mount. Dnotify is therefore infeasible
  134. for use on a desktop system with removable media as the media cannot be
  135. unmounted. Watching a file should not require that it be open.
  136. Q: What is the design decision behind using an-fd-per-instance as opposed to
  137. an fd-per-watch?
  138. A: An fd-per-watch quickly consumes more file descriptors than are allowed,
  139. more fd's than are feasible to manage, and more fd's than are optimally
  140. select()-able. Yes, root can bump the per-process fd limit and yes, users
  141. can use epoll, but requiring both is a silly and extraneous requirement.
  142. A watch consumes less memory than an open file, separating the number
  143. spaces is thus sensible. The current design is what user-space developers
  144. want: Users initialize inotify, once, and add n watches, requiring but one
  145. fd and no twiddling with fd limits. Initializing an inotify instance two
  146. thousand times is silly. If we can implement user-space's preferences
  147. cleanly--and we can, the idr layer makes stuff like this trivial--then we
  148. should.
  149. There are other good arguments. With a single fd, there is a single
  150. item to block on, which is mapped to a single queue of events. The single
  151. fd returns all watch events and also any potential out-of-band data. If
  152. every fd was a separate watch,
  153. - There would be no way to get event ordering. Events on file foo and
  154. file bar would pop poll() on both fd's, but there would be no way to tell
  155. which happened first. A single queue trivially gives you ordering. Such
  156. ordering is crucial to existing applications such as Beagle. Imagine
  157. "mv a b ; mv b a" events without ordering.
  158. - We'd have to maintain n fd's and n internal queues with state,
  159. versus just one. It is a lot messier in the kernel. A single, linear
  160. queue is the data structure that makes sense.
  161. - User-space developers prefer the current API. The Beagle guys, for
  162. example, love it. Trust me, I asked. It is not a surprise: Who'd want
  163. to manage and block on 1000 fd's via select?
  164. - No way to get out of band data.
  165. - 1024 is still too low. ;-)
  166. When you talk about designing a file change notification system that
  167. scales to 1000s of directories, juggling 1000s of fd's just does not seem
  168. the right interface. It is too heavy.
  169. Additionally, it _is_ possible to more than one instance and
  170. juggle more than one queue and thus more than one associated fd. There
  171. need not be a one-fd-per-process mapping; it is one-fd-per-queue and a
  172. process can easily want more than one queue.
  173. Q: Why the system call approach?
  174. A: The poor user-space interface is the second biggest problem with dnotify.
  175. Signals are a terrible, terrible interface for file notification. Or for
  176. anything, for that matter. The ideal solution, from all perspectives, is a
  177. file descriptor-based one that allows basic file I/O and poll/select.
  178. Obtaining the fd and managing the watches could have been done either via a
  179. device file or a family of new system calls. We decided to implement a
  180. family of system calls because that is the preferred approach for new kernel
  181. interfaces. The only real difference was whether we wanted to use open(2)
  182. and ioctl(2) or a couple of new system calls. System calls beat ioctls.