91 lines · plain
1.. SPDX-License-Identifier: GPL-2.02 3===============================================================4Inotify - A Powerful yet Simple File Change Notification System5===============================================================6 7 8 9Document started 15 Mar 2005 by Robert Love <rml@novell.com>10 11Document updated 4 Jan 2015 by Zhang Zhen <zhenzhang.zhang@huawei.com>12 13 - Deleted obsoleted interface, just refer to manpages for user interface.14 15(i) Rationale16 17Q:18 What is the design decision behind not tying the watch to the open fd of19 the watched object?20 21A:22 Watches are associated with an open inotify device, not an open file.23 This solves the primary problem with dnotify: keeping the file open pins24 the file and thus, worse, pins the mount. Dnotify is therefore infeasible25 for use on a desktop system with removable media as the media cannot be26 unmounted. Watching a file should not require that it be open.27 28Q:29 What is the design decision behind using an-fd-per-instance as opposed to30 an fd-per-watch?31 32A:33 An fd-per-watch quickly consumes more file descriptors than are allowed,34 more fd's than are feasible to manage, and more fd's than are optimally35 select()-able. Yes, root can bump the per-process fd limit and yes, users36 can use epoll, but requiring both is a silly and extraneous requirement.37 A watch consumes less memory than an open file, separating the number38 spaces is thus sensible. The current design is what user-space developers39 want: Users initialize inotify, once, and add n watches, requiring but one40 fd and no twiddling with fd limits. Initializing an inotify instance two41 thousand times is silly. If we can implement user-space's preferences42 cleanly--and we can, the idr layer makes stuff like this trivial--then we43 should.44 45 There are other good arguments. With a single fd, there is a single46 item to block on, which is mapped to a single queue of events. The single47 fd returns all watch events and also any potential out-of-band data. If48 every fd was a separate watch,49 50 - There would be no way to get event ordering. Events on file foo and51 file bar would pop poll() on both fd's, but there would be no way to tell52 which happened first. A single queue trivially gives you ordering. Such53 ordering is crucial to existing applications such as Beagle. Imagine54 "mv a b ; mv b a" events without ordering.55 56 - We'd have to maintain n fd's and n internal queues with state,57 versus just one. It is a lot messier in the kernel. A single, linear58 queue is the data structure that makes sense.59 60 - User-space developers prefer the current API. The Beagle guys, for61 example, love it. Trust me, I asked. It is not a surprise: Who'd want62 to manage and block on 1000 fd's via select?63 64 - No way to get out of band data.65 66 - 1024 is still too low. ;-)67 68 When you talk about designing a file change notification system that69 scales to 1000s of directories, juggling 1000s of fd's just does not seem70 the right interface. It is too heavy.71 72 Additionally, it _is_ possible to more than one instance and73 juggle more than one queue and thus more than one associated fd. There74 need not be a one-fd-per-process mapping; it is one-fd-per-queue and a75 process can easily want more than one queue.76 77Q:78 Why the system call approach?79 80A:81 The poor user-space interface is the second biggest problem with dnotify.82 Signals are a terrible, terrible interface for file notification. Or for83 anything, for that matter. The ideal solution, from all perspectives, is a84 file descriptor-based one that allows basic file I/O and poll/select.85 Obtaining the fd and managing the watches could have been done either via a86 device file or a family of new system calls. We decided to implement a87 family of system calls because that is the preferred approach for new kernel88 interfaces. The only real difference was whether we wanted to use open(2)89 and ioctl(2) or a couple of new system calls. System calls beat ioctls.90 91