NetBSD/dist/am-utils/README.autofs

# -*- text -*-
# This file is README.autofs
# am-utils-6.1
# Erez Zadok <ezk@cs.columbia.edu> a.k.a. "The Lord of Darkness"
#
# modified by
# Ion Badulescu <ib42@cs.columbia.edu> a.k.a. "The Autofs Master"

** General notes about the autofs support in am-utils

- The autofs code in am-utils is gamma quality for Linux, beta quality for
  Solaris 2.5+ and non-working for all the other systems.

- Link, lofs, and nfs mounts were tested and work properly; the others
  should work, but were not tested.

* Caveats:

- [this applies to Solaris/Irix/HP-UX/AIX] Amd acts as *both* automountd and
automount.  There's no way to distinguish between the two.  When amd starts,
it first registers itself as an autofs server (automountd's job), then
parses its own maps, and decides which autofs-type mounts to make
(automount's job).  After the autofs mounts are made, amd listens for
requests from the kernel-based autofs, and acts upon them. Since there can
be only one autofs listener on a system, this means that automountd and amd
cannot run at the same time; nor can two amd's run at the same time if they
are both using autofs mounts.

- Linux support is available and fairly stable. Solaris 2.5+ support is
newer and less tested, but seems pretty stable as well. Irix and HP-UX
autofs support will probably be very easy once we get Solaris 2.5 to work,
as they use the same protocol and almost identical data structures. AIX
seems to be using the Solaris 2.5 protocol as well, but they don't provide
any headers or documentation so getting autofs to work will be tricky at best.

- Killing amd can become a problem if there are active mounts. Since mounts
are done "in place", we can't just unmount our mount points and go away. For
now, amd simply tells the kernel that it is dying; we need to think this
further.  It would be nice to "take over" the old mountpoints, there is
support for this in Solaris (due to is RPC nature) and is easy to add to
the Linux kernel (an ioctl on the mountpoint).

- The Solaris 2.5 autofs protocol is NON-REENTRANT by design, and ignoring
this limitation results in a DEADLOCK between the kernel and the
daemon. This is a serious problem, although only for the lofs and link mount
types. In other words, if the daemon is trying a lofs/link mount whose
destination crosses an autofs mountpoint, it will trigger another autofs
lookup which will deadlock inside the kernel -- because the kernel is
waiting for the previous request to return. This is not even something
specific to amd, Sun's own automountd has the exact same problem. Ctrl-C
breaks the deadlock, so it's not fatal, but the lofs/link entry is in effect
unusable.

* Solaris:

- Amd w/ autofs mounts will fight over the listener port with Sun's
  automountd, so running both simultaneously is a really bad idea.

- Browsable_dirs is possible and implemented for Solaris 2.6+.

- Direct mounts are implemented and work correctly.

- Link mounts are implemented as symlinks on Solaris 2.6+, but are
  inefficient. The kernel seems to time them out immediately after receiving
  them, so each access to a symlink causes a call to userspace. Needless to
  say, this negates the whole point of using autofs. Automountd seems to
  always use lofs mounts instead of symlinks, we should probably do the same.

- Link mounts are implemented as lofs mounts on Solaris 2.5/2.5.1, subject to
  the limitation described in the caveats section above.

- Restarting autofs mounts is possible, but not yet implemented.

* Linux:

- Amd should work fine even when the Linux automounter is running, the
  mechanism being used prevents any kind of (evil) interaction between them.

- Browsing is not available if autofs support is used, due to limitations in
  the kernel-daemon protocol used by Linux 2.2 and 2.4. Only already-mounted
  nodes will appear in the autofs directory, and this is implemented entirely
  in the kernel.

- Host maps are supported with autofs4 (in Linux 2.3+). You need to add
  something like "alias autofs=autofs4" in /etc/modules.conf.

- Direct maps cannot be supported since there is no kernel support for them;
  we might be able to get something eventually, but don't hold your breath.
  If anything, we may do it using a regular NFS mountpoint and bind-mount on
  top of it.

- Inherit doesn't make much sense because we can't restart a hung autofs
  mount point, due to kernel limitations. This needs to be fixed in the
  Linux kernel; it's not particularly difficult, and we might provide a
  patch at some point.

- Link (and lofs) mounts will use the new bind-mount support in Linux
  2.4+. No more symlinks! And /bin/pwd works great too.

* amd.conf requirements:

To tell amd to use an autofs-style mount point and mounts for a map, add

	mount_type = autofs

either to the global section, or to the sections of the individual maps you
selected. Mixing autofs and normal amd mount points in the same amd.conf
file *does* work.

* Map changes:

No changes are necessary. If a map is marked as autofs in amd.conf, mounts
are done "in place" and the "fs" parameter is ignored most of the time.

* Todo:

We are looking for volunteers to improve the autofs code!

(1) These fixes are needed:

- when the mount type is 'link', transparently translate it into a loopback
file system mount (lofs), that would mount in place, rather than supply a
symlink, at least on systems whose autofs doesn't support symlinks. Linux
does support symlinks, Solaris 2.6+ does too, but Solaris 2.5/2.5.1 doesn't
and neither does Irix. Moreover, Sun's automountd always uses lofs for
link mounts, even on 2.6+, because symlinks are not cached in the kernel and
thus are not particularly efficient. [already done for Linux 2.4+ using bind
mounts and for Solaris 2.5 using lofs mounts]

- complain if certain incompatible options (autofs and ...) are used. Direct
maps on Linux is one such case of incompatible options. browsable_dirs on
Linux is another such case.

- if amd is killed or dies, the autofs mounts will remain intact as required
(your system is _not_ hung, yay!).  However, if you restart amd, it will not
correctly 'restart' the autofs mounts as the Sun automounter does.  Rather,
it might cause another mount to happen, which leaves your /etc/mnttab
cluttered with older mounts entries which cannot be unmounted. It might also
just pretend everything is ok, when in fact it isn't.

(2) Code expansion:

- [Solaris only] implement the sockets version of amu_get_autofs_address()
and create_autofs_service(), in conf/transp/transp_sockets.c. Not sure if
it's necessary, Solaris it still biased towards TLI/STREAMS in userspace.

- Implement the restarting of autofs mount points. This is already doable on
Solaris; on Linux, the kernel needs to be patched to allow it.

(3) Testing and porting to other systems:

- nothing has been tested on Irix, which reportedly has a similarly
functioning autofs to Solaris 2.5.

- support for Linux autofs is stable, we need testers!

- support for Solaris 2.6+ is pretty stable, so we need testers for it, too!

- we did not test any version of Solaris on x86. It will probably work, but
you have been warned. Testers are welcome.