blob: 8fac3fe7b8c971c0c39e54283c2fa0698922aaff [file] [log] [blame]
NeilBrown87d672c2014-10-13 15:52:24 -07001<head>
2<style> p { max-width:50em} ol, ul {max-width: 40em}</style>
3</head>
4
5autofs - how it works
6=====================
7
8Purpose
9-------
10
11The goal of autofs is to provide on-demand mounting and race free
12automatic unmounting of various other filesystems. This provides two
13key advantages:
14
151. There is no need to delay boot until all filesystems that
16 might be needed are mounted. Processes that try to access those
17 slow filesystems might be delayed but other processes can
18 continue freely. This is particularly important for
19 network filesystems (e.g. NFS) or filesystems stored on
20 media with a media-changing robot.
21
222. The names and locations of filesystems can be stored in
23 a remote database and can change at any time. The content
24 in that data base at the time of access will be used to provide
25 a target for the access. The interpretation of names in the
26 filesystem can even be programmatic rather than database-backed,
27 allowing wildcards for example, and can vary based on the user who
28 first accessed a name.
29
30Context
31-------
32
33The "autofs4" filesystem module is only one part of an autofs system.
34There also needs to be a user-space program which looks up names
35and mounts filesystems. This will often be the "automount" program,
36though other tools including "systemd" can make use of "autofs4".
37This document describes only the kernel module and the interactions
38required with any user-space program. Subsequent text refers to this
39as the "automount daemon" or simply "the daemon".
40
41"autofs4" is a Linux kernel module with provides the "autofs"
42filesystem type. Several "autofs" filesystems can be mounted and they
43can each be managed separately, or all managed by the same daemon.
44
45Content
46-------
47
48An autofs filesystem can contain 3 sorts of objects: directories,
49symbolic links and mount traps. Mount traps are directories with
50extra properties as described in the next section.
51
52Objects can only be created by the automount daemon: symlinks are
53created with a regular `symlink` system call, while directories and
54mount traps are created with `mkdir`. The determination of whether a
55directory should be a mount trap or not is quite _ad hoc_, largely for
56historical reasons, and is determined in part by the
57*direct*/*indirect*/*offset* mount options, and the *maxproto* mount option.
58
59If neither the *direct* or *offset* mount options are given (so the
60mount is considered to be *indirect*), then the root directory is
61always a regular directory, otherwise it is a mount trap when it is
62empty and a regular directory when not empty. Note that *direct* and
63*offset* are treated identically so a concise summary is that the root
64directory is a mount trap only if the filesystem is mounted *direct*
65and the root is empty.
66
67Directories created in the root directory are mount traps only if the
68filesystem is mounted *indirect* and they are empty.
69
70Directories further down the tree depend on the *maxproto* mount
71option and particularly whether it is less than five or not.
72When *maxproto* is five, no directories further down the
73tree are ever mount traps, they are always regular directories. When
74the *maxproto* is four (or three), these directories are mount traps
75precisely when they are empty.
76
77So: non-empty (i.e. non-leaf) directories are never mount traps. Empty
78directories are sometimes mount traps, and sometimes not depending on
79where in the tree they are (root, top level, or lower), the *maxproto*,
80and whether the mount was *indirect* or not.
81
82Mount Traps
83---------------
84
85A core element of the implementation of autofs is the Mount Traps
86which are provided by the Linux VFS. Any directory provided by a
87filesystem can be designated as a trap. This involves two separate
88features that work together to allow autofs to do its job.
89
90**DCACHE_NEED_AUTOMOUNT**
91
92If a dentry has the DCACHE_NEED_AUTOMOUNT flag set (which gets set if
93the inode has S_AUTOMOUNT set, or can be set directly) then it is
94(potentially) a mount trap. Any access to this directory beyond a
95"`stat`" will (normally) cause the `d_op->d_automount()` dentry operation
96to be called. The task of this method is to find the filesystem that
97should be mounted on the directory and to return it. The VFS is
98responsible for actually mounting the root of this filesystem on the
99directory.
100
101autofs doesn't find the filesystem itself but sends a message to the
102automount daemon asking it to find and mount the filesystem. The
103autofs `d_automount` method then waits for the daemon to report that
104everything is ready. It will then return "`NULL`" indicating that the
105mount has already happened. The VFS doesn't try to mount anything but
106follows down the mount that is already there.
107
108This functionality is sufficient for some users of mount traps such
109as NFS which creates traps so that mountpoints on the server can be
110reflected on the client. However it is not sufficient for autofs. As
111mounting onto a directory is considered to be "beyond a `stat`", the
112automount daemon would not be able to mount a filesystem on the 'trap'
113directory without some way to avoid getting caught in the trap. For
114that purpose there is another flag.
115
116**DCACHE_MANAGE_TRANSIT**
117
118If a dentry has DCACHE_MANAGE_TRANSIT set then two very different but
119related behaviors are invoked, both using the `d_op->d_manage()`
120dentry operation.
121
122Firstly, before checking to see if any filesystem is mounted on the
123directory, d_manage() will be called with the `rcu_walk` parameter set
124to `false`. It may return one of three things:
125
126- A return value of zero indicates that there is nothing special
127 about this dentry and normal checks for mounts and automounts
128 should proceed.
129
130 autofs normally returns zero, but first waits for any
131 expiry (automatic unmounting of the mounted filesystem) to
132 complete. This avoids races.
133
134- A return value of `-EISDIR` tells the VFS to ignore any mounts
135 on the directory and to not consider calling `->d_automount()`.
136 This effectively disables the **DCACHE_NEED_AUTOMOUNT** flag
137 causing the directory not be a mount trap after all.
138
139 autofs returns this if it detects that the process performing the
140 lookup is the automount daemon and that the mount has been
141 requested but has not yet completed. How it determines this is
142 discussed later. This allows the automount daemon not to get
143 caught in the mount trap.
144
145 There is a subtlety here. It is possible that a second autofs
146 filesystem can be mounted below the first and for both of them to
147 be managed by the same daemon. For the daemon to be able to mount
148 something on the second it must be able to "walk" down past the
149 first. This means that d_manage cannot *always* return -EISDIR for
150 the automount daemon. It must only return it when a mount has
151 been requested, but has not yet completed.
152
153 `d_manage` also returns `-EISDIR` if the dentry shouldn't be a
154 mount trap, either because it is a symbolic link or because it is
155 not empty.
156
157- Any other negative value is treated as an error and returned
158 to the caller.
159
160 autofs can return
161
162 - -ENOENT if the automount daemon failed to mount anything,
163 - -ENOMEM if it ran out of memory,
164 - -EINTR if a signal arrived while waiting for expiry to
165 complete
166 - or any other error sent down by the automount daemon.
167
168
169The second use case only occurs during an "RCU-walk" and so `rcu_walk`
170will be set.
171
172An RCU-walk is a fast and lightweight process for walking down a
173filename path (i.e. it is like running on tip-toes). RCU-walk cannot
174cope with all situations so when it finds a difficulty it falls back
175to "REF-walk", which is slower but more robust.
176
177RCU-walk will never call `->d_automount`; the filesystems must already
178be mounted or RCU-walk cannot handle the path.
179To determine if a mount-trap is safe for RCU-walk mode it calls
180`->d_manage()` with `rcu_walk` set to `true`.
181
182In this case `d_manage()` must avoid blocking and should avoid taking
183spinlocks if at all possible. Its sole purpose is to determine if it
184would be safe to follow down into any mounted directory and the only
185reason that it might not be is if an expiry of the mount is
186underway.
187
188In the `rcu_walk` case, `d_manage()` cannot return -EISDIR to tell the
189VFS that this is a directory that doesn't require d_automount. If
190`rcu_walk` sees a dentry with DCACHE_NEED_AUTOMOUNT set but nothing
191mounted, it *will* fall back to REF-walk. `d_manage()` cannot make the
192VFS remain in RCU-walk mode, but can only tell it to get out of
193RCU-walk mode by returning `-ECHILD`.
194
195So `d_manage()`, when called with `rcu_walk` set, should either return
196-ECHILD if there is any reason to believe it is unsafe to end the
197mounted filesystem, and otherwise should return 0.
198
199autofs will return `-ECHILD` if an expiry of the filesystem has been
200initiated or is being considered, otherwise it returns 0.
201
202
203Mountpoint expiry
204-----------------
205
Tomohiro Kusumie6621452016-10-11 13:52:25 -0700206The VFS has a mechanism for automatically expiring unused mounts,
NeilBrown87d672c2014-10-13 15:52:24 -0700207much as it can expire any unused dentry information from the dcache.
Tomohiro Kusumie6621452016-10-11 13:52:25 -0700208This is guided by the MNT_SHRINKABLE flag. This only applies to
NeilBrown87d672c2014-10-13 15:52:24 -0700209mounts that were created by `d_automount()` returning a filesystem to be
210mounted. As autofs doesn't return such a filesystem but leaves the
211mounting to the automount daemon, it must involve the automount daemon
212in unmounting as well. This also means that autofs has more control
213of expiry.
214
215The VFS also supports "expiry" of mounts using the MNT_EXPIRE flag to
216the `umount` system call. Unmounting with MNT_EXPIRE will fail unless
217a previous attempt had been made, and the filesystem has been inactive
218and untouched since that previous attempt. autofs4 does not depend on
219this but has its own internal tracking of whether filesystems were
220recently used. This allows individual names in the autofs directory
221to expire separately.
222
223With version 4 of the protocol, the automount daemon can try to
224unmount any filesystems mounted on the autofs filesystem or remove any
225symbolic links or empty directories any time it likes. If the unmount
226or removal is successful the filesystem will be returned to the state
227it was before the mount or creation, so that any access of the name
228will trigger normal auto-mount processing. In particlar, `rmdir` and
229`unlink` do not leave negative entries in the dcache as a normal
230filesystem would, so an attempt to access a recently-removed object is
231passed to autofs for handling.
232
233With version 5, this is not safe except for unmounting from top-level
234directories. As lower-level directories are never mount traps, other
235processes will see an empty directory as soon as the filesystem is
236unmounted. So it is generally safest to use the autofs expiry
237protocol described below.
238
239Normally the daemon only wants to remove entries which haven't been
240used for a while. For this purpose autofs maintains a "`last_used`"
241time stamp on each directory or symlink. For symlinks it genuinely
242does record the last time the symlink was "used" or followed to find
243out where it points to. For directories the field is a slight
244misnomer. It actually records the last time that autofs checked if
245the directory or one of its descendents was busy and found that it
246was. This is just as useful and doesn't require updating the field so
247often.
248
249The daemon is able to ask autofs if anything is due to be expired,
250using an `ioctl` as discussed later. For a *direct* mount, autofs
251considers if the entire mount-tree can be unmounted or not. For an
252*indirect* mount, autofs considers each of the names in the top level
253directory to determine if any of those can be unmounted and cleaned
254up.
255
256There is an option with indirect mounts to consider each of the leaves
257that has been mounted on instead of considering the top-level names.
258This is intended for compatability with version 4 of autofs and should
259be considered as deprecated.
260
261When autofs considers a directory it checks the `last_used` time and
262compares it with the "timeout" value set when the filesystem was
263mounted, though this check is ignored in some cases. It also checks if
264the directory or anything below it is in use. For symbolic links,
265only the `last_used` time is ever considered.
266
267If both appear to support expiring the directory or symlink, an action
268is taken.
269
270There are two ways to ask autofs to consider expiry. The first is to
271use the **AUTOFS_IOC_EXPIRE** ioctl. This only works for indirect
272mounts. If it finds something in the root directory to expire it will
273return the name of that thing. Once a name has been returned the
274automount daemon needs to unmount any filesystems mounted below the
275name normally. As described above, this is unsafe for non-toplevel
276mounts in a version-5 autofs. For this reason the current `automountd`
277does not use this ioctl.
278
279The second mechanism uses either the **AUTOFS_DEV_IOCTL_EXPIRE_CMD** or
280the **AUTOFS_IOC_EXPIRE_MULTI** ioctl. This will work for both direct and
281indirect mounts. If it selects an object to expire, it will notify
282the daemon using the notification mechanism described below. This
283will block until the daemon acknowledges the expiry notification.
284This implies that the "`EXPIRE`" ioctl must be sent from a different
285thread than the one which handles notification.
286
287While the ioctl is blocking, the entry is marked as "expiring" and
288`d_manage` will block until the daemon affirms that the unmount has
289completed (together with removing any directories that might have been
290necessary), or has been aborted.
291
292Communicating with autofs: detecting the daemon
293-----------------------------------------------
294
295There are several forms of communication between the automount daemon
296and the filesystem. As we have already seen, the daemon can create and
297remove directories and symlinks using normal filesystem operations.
298autofs knows whether a process requesting some operation is the daemon
299or not based on its process-group id number (see getpgid(1)).
300
Tomohiro Kusumie6621452016-10-11 13:52:25 -0700301When an autofs filesystem is mounted the pgid of the mounting
NeilBrown87d672c2014-10-13 15:52:24 -0700302processes is recorded unless the "pgrp=" option is given, in which
303case that number is recorded instead. Any request arriving from a
304process in that process group is considered to come from the daemon.
305If the daemon ever has to be stopped and restarted a new pgid can be
306provided through an ioctl as will be described below.
307
308Communicating with autofs: the event pipe
309-----------------------------------------
310
311When an autofs filesystem is mounted, the 'write' end of a pipe must
312be passed using the 'fd=' mount option. autofs will write
313notification messages to this pipe for the daemon to respond to.
314For version 5, the format of the message is:
315
316 struct autofs_v5_packet {
317 int proto_version; /* Protocol version */
318 int type; /* Type of packet */
319 autofs_wqt_t wait_queue_token;
320 __u32 dev;
321 __u64 ino;
322 __u32 uid;
323 __u32 gid;
324 __u32 pid;
325 __u32 tgid;
326 __u32 len;
327 char name[NAME_MAX+1];
328 };
329
330where the type is one of
331
332 autofs_ptype_missing_indirect
333 autofs_ptype_expire_indirect
334 autofs_ptype_missing_direct
335 autofs_ptype_expire_direct
336
337so messages can indicate that a name is missing (something tried to
338access it but it isn't there) or that it has been selected for expiry.
339
340The pipe will be set to "packet mode" (equivalent to passing
341`O_DIRECT`) to _pipe2(2)_ so that a read from the pipe will return at
342most one packet, and any unread portion of a packet will be discarded.
343
344The `wait_queue_token` is a unique number which can identify a
345particular request to be acknowledged. When a message is sent over
346the pipe the affected dentry is marked as either "active" or
347"expiring" and other accesses to it block until the message is
348acknowledged using one of the ioctls below and the relevant
349`wait_queue_token`.
350
351Communicating with autofs: root directory ioctls
352------------------------------------------------
353
354The root directory of an autofs filesystem will respond to a number of
355ioctls. The process issuing the ioctl must have the CAP_SYS_ADMIN
356capability, or must be the automount daemon.
357
358The available ioctl commands are:
359
360- **AUTOFS_IOC_READY**: a notification has been handled. The argument
361 to the ioctl command is the "wait_queue_token" number
362 corresponding to the notification being acknowledged.
363- **AUTOFS_IOC_FAIL**: similar to above, but indicates failure with
364 the error code `ENOENT`.
365- **AUTOFS_IOC_CATATONIC**: Causes the autofs to enter "catatonic"
366 mode meaning that it stops sending notifications to the daemon.
367 This mode is also entered if a write to the pipe fails.
368- **AUTOFS_IOC_PROTOVER**: This returns the protocol version in use.
369- **AUTOFS_IOC_PROTOSUBVER**: Returns the protocol sub-version which
370 is really a version number for the implementation. It is
371 currently 2.
372- **AUTOFS_IOC_SETTIMEOUT**: This passes a pointer to an unsigned
373 long. The value is used to set the timeout for expiry, and
374 the current timeout value is stored back through the pointer.
375- **AUTOFS_IOC_ASKUMOUNT**: Returns, in the pointed-to `int`, 1 if
376 the filesystem could be unmounted. This is only a hint as
377 the situation could change at any instant. This call can be
378 use to avoid a more expensive full unmount attempt.
379- **AUTOFS_IOC_EXPIRE**: as described above, this asks if there is
380 anything suitable to expire. A pointer to a packet:
381
382 struct autofs_packet_expire_multi {
383 int proto_version; /* Protocol version */
384 int type; /* Type of packet */
385 autofs_wqt_t wait_queue_token;
386 int len;
387 char name[NAME_MAX+1];
388 };
389
390 is required. This is filled in with the name of something
391 that can be unmounted or removed. If nothing can be expired,
392 `errno` is set to `EAGAIN`. Even though a `wait_queue_token`
393 is present in the structure, no "wait queue" is established
394 and no acknowledgment is needed.
395- **AUTOFS_IOC_EXPIRE_MULTI**: This is similar to
396 **AUTOFS_IOC_EXPIRE** except that it causes notification to be
397 sent to the daemon, and it blocks until the daemon acknowledges.
398 The argument is an integer which can contain two different flags.
399
400 **AUTOFS_EXP_IMMEDIATE** causes `last_used` time to be ignored
401 and objects are expired if the are not in use.
402
403 **AUTOFS_EXP_LEAVES** will select a leaf rather than a top-level
404 name to expire. This is only safe when *maxproto* is 4.
405
406Communicating with autofs: char-device ioctls
407---------------------------------------------
408
409It is not always possible to open the root of an autofs filesystem,
410particularly a *direct* mounted filesystem. If the automount daemon
411is restarted there is no way for it to regain control of existing
412mounts using any of the above communication channels. To address this
413need there is a "miscellaneous" character device (major 10, minor 235)
414which can be used to communicate directly with the autofs filesystem.
415It requires CAP_SYS_ADMIN for access.
416
417The `ioctl`s that can be used on this device are described in a separate
418document `autofs4-mount-control.txt`, and are summarized briefly here.
419Each ioctl is passed a pointer to an `autofs_dev_ioctl` structure:
420
421 struct autofs_dev_ioctl {
422 __u32 ver_major;
423 __u32 ver_minor;
424 __u32 size; /* total size of data passed in
425 * including this struct */
426 __s32 ioctlfd; /* automount command fd */
427
428 __u32 arg1; /* Command parameters */
429 __u32 arg2;
430
431 char path[0];
432 };
433
434For the **OPEN_MOUNT** and **IS_MOUNTPOINT** commands, the target
435filesystem is identified by the `path`. All other commands identify
436the filesystem by the `ioctlfd` which is a file descriptor open on the
437root, and which can be returned by **OPEN_MOUNT**.
438
439The `ver_major` and `ver_minor` are in/out parameters which check that
440the requested version is supported, and report the maximum version
441that the kernel module can support.
442
443Commands are:
444
445- **AUTOFS_DEV_IOCTL_VERSION_CMD**: does nothing, except validate and
446 set version numbers.
447- **AUTOFS_DEV_IOCTL_OPENMOUNT_CMD**: return an open file descriptor
448 on the root of an autofs filesystem. The filesystem is identified
449 by name and device number, which is stored in `arg1`. Device
450 numbers for existing filesystems can be found in
451 `/proc/self/mountinfo`.
452- **AUTOFS_DEV_IOCTL_CLOSEMOUNT_CMD**: same as `close(ioctlfd)`.
Tomohiro Kusumie6621452016-10-11 13:52:25 -0700453- **AUTOFS_DEV_IOCTL_SETPIPEFD_CMD**: if the filesystem is in
NeilBrown87d672c2014-10-13 15:52:24 -0700454 catatonic mode, this can provide the write end of a new pipe
455 in `arg1` to re-establish communication with a daemon. The
456 process group of the calling process is used to identify the
457 daemon.
458- **AUTOFS_DEV_IOCTL_REQUESTER_CMD**: `path` should be a
459 name within the filesystem that has been auto-mounted on.
460 arg1 is the dev number of the underlying autofs. On successful
461 return, `arg1` and `arg2` will be the UID and GID of the process
462 which triggered that mount.
463
464- **AUTOFS_DEV_IOCTL_ISMOUNTPOINT_CMD**: Check if path is a
465 mountpoint of a particular type - see separate documentation for
466 details.
467
468- **AUTOFS_DEV_IOCTL_PROTOVER_CMD**:
469- **AUTOFS_DEV_IOCTL_PROTOSUBVER_CMD**:
470- **AUTOFS_DEV_IOCTL_READY_CMD**:
471- **AUTOFS_DEV_IOCTL_FAIL_CMD**:
472- **AUTOFS_DEV_IOCTL_CATATONIC_CMD**:
473- **AUTOFS_DEV_IOCTL_TIMEOUT_CMD**:
474- **AUTOFS_DEV_IOCTL_EXPIRE_CMD**:
475- **AUTOFS_DEV_IOCTL_ASKUMOUNT_CMD**: These all have the same
476 function as the similarly named **AUTOFS_IOC** ioctls, except
477 that **FAIL** can be given an explicit error number in `arg1`
478 instead of assuming `ENOENT`, and this **EXPIRE** command
479 corresponds to **AUTOFS_IOC_EXPIRE_MULTI**.
480
481Catatonic mode
482--------------
483
484As mentioned, an autofs mount can enter "catatonic" mode. This
485happens if a write to the notification pipe fails, or if it is
486explicitly requested by an `ioctl`.
487
488When entering catatonic mode, the pipe is closed and any pending
489notifications are acknowledged with the error `ENOENT`.
490
491Once in catatonic mode attempts to access non-existing names will
492result in `ENOENT` while attempts to access existing directories will
493be treated in the same way as if they came from the daemon, so mount
494traps will not fire.
495
496When the filesystem is mounted a _uid_ and _gid_ can be given which
497set the ownership of directories and symbolic links. When the
498filesystem is in catatonic mode, any process with a matching UID can
499create directories or symlinks in the root directory, but not in other
500directories.
501
502Catatonic mode can only be left via the
503**AUTOFS_DEV_IOCTL_OPENMOUNT_CMD** ioctl on the `/dev/autofs`.
504
505autofs, name spaces, and shared mounts
506--------------------------------------
507
508With bind mounts and name spaces it is possible for an autofs
509filesystem to appear at multiple places in one or more filesystem
510name spaces. For this to work sensibly, the autofs filesystem should
511always be mounted "shared". e.g.
512
513> `mount --make-shared /autofs/mount/point`
514
515The automount daemon is only able to mange a single mount location for
516an autofs filesystem and if mounts on that are not 'shared', other
517locations will not behave as expected. In particular access to those
518other locations will likely result in the `ELOOP` error
519
520> Too many levels of symbolic links