inotify.rst 4.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990
  1. .. SPDX-License-Identifier: GPL-2.0
  2. ===============================================================
  3. Inotify - A Powerful yet Simple File Change Notification System
  4. ===============================================================
  5. Document started 15 Mar 2005 by Robert Love <rml@novell.com>
  6. Document updated 4 Jan 2015 by Zhang Zhen <zhenzhang.zhang@huawei.com>
  7. - Deleted obsoleted interface, just refer to manpages for user interface.
  8. (i) Rationale
  9. Q:
  10. What is the design decision behind not tying the watch to the open fd of
  11. the watched object?
  12. A:
  13. Watches are associated with an open inotify device, not an open file.
  14. This solves the primary problem with dnotify: keeping the file open pins
  15. the file and thus, worse, pins the mount. Dnotify is therefore infeasible
  16. for use on a desktop system with removable media as the media cannot be
  17. unmounted. Watching a file should not require that it be open.
  18. Q:
  19. What is the design decision behind using an-fd-per-instance as opposed to
  20. an fd-per-watch?
  21. A:
  22. An fd-per-watch quickly consumes more file descriptors than are allowed,
  23. more fd's than are feasible to manage, and more fd's than are optimally
  24. select()-able. Yes, root can bump the per-process fd limit and yes, users
  25. can use epoll, but requiring both is a silly and extraneous requirement.
  26. A watch consumes less memory than an open file, separating the number
  27. spaces is thus sensible. The current design is what user-space developers
  28. want: Users initialize inotify, once, and add n watches, requiring but one
  29. fd and no twiddling with fd limits. Initializing an inotify instance two
  30. thousand times is silly. If we can implement user-space's preferences
  31. cleanly--and we can, the idr layer makes stuff like this trivial--then we
  32. should.
  33. There are other good arguments. With a single fd, there is a single
  34. item to block on, which is mapped to a single queue of events. The single
  35. fd returns all watch events and also any potential out-of-band data. If
  36. every fd was a separate watch,
  37. - There would be no way to get event ordering. Events on file foo and
  38. file bar would pop poll() on both fd's, but there would be no way to tell
  39. which happened first. A single queue trivially gives you ordering. Such
  40. ordering is crucial to existing applications such as Beagle. Imagine
  41. "mv a b ; mv b a" events without ordering.
  42. - We'd have to maintain n fd's and n internal queues with state,
  43. versus just one. It is a lot messier in the kernel. A single, linear
  44. queue is the data structure that makes sense.
  45. - User-space developers prefer the current API. The Beagle guys, for
  46. example, love it. Trust me, I asked. It is not a surprise: Who'd want
  47. to manage and block on 1000 fd's via select?
  48. - No way to get out of band data.
  49. - 1024 is still too low. ;-)
  50. When you talk about designing a file change notification system that
  51. scales to 1000s of directories, juggling 1000s of fd's just does not seem
  52. the right interface. It is too heavy.
  53. Additionally, it _is_ possible to more than one instance and
  54. juggle more than one queue and thus more than one associated fd. There
  55. need not be a one-fd-per-process mapping; it is one-fd-per-queue and a
  56. process can easily want more than one queue.
  57. Q:
  58. Why the system call approach?
  59. A:
  60. The poor user-space interface is the second biggest problem with dnotify.
  61. Signals are a terrible, terrible interface for file notification. Or for
  62. anything, for that matter. The ideal solution, from all perspectives, is a
  63. file descriptor-based one that allows basic file I/O and poll/select.
  64. Obtaining the fd and managing the watches could have been done either via a
  65. device file or a family of new system calls. We decided to implement a
  66. family of system calls because that is the preferred approach for new kernel
  67. interfaces. The only real difference was whether we wanted to use open(2)
  68. and ioctl(2) or a couple of new system calls. System calls beat ioctls.