TEMPLATING machine configurations
Matthew Dillon
dillon@backplane.com
This document describes a general mechanism by which you can template
/ and /usr. That is, to keep a 'master template' of / and /usr on a
separate machine which is then used to update the rest of your machines.
Generally speaking, you can't simply mirror /. You might be able to
get away with mirroring /usr. There are two main problems involved with
templating:
(1) Avoiding overwriting run-time generated files
By default, the system maintains a number of files in the root
partition. For example, sendmail will dbm /etc/aliases into
/etc/aliases.db. vipw or chpass or other password related routines
will regenerate the password dbm's /etc/spwd.db, /etc/pwd.db, and
passwd. /etc/namedb/s might contain generated secondaries. And
so forth.
The templating mechanism must avoid copying over such files.
(2) Customizing machines.
Customizing machines is actually considerably simpler. You create
a configuration hierarchy and convert the configuration files that
have to be customized into softlinks that run through a special
softlink in the configuration directory. This will work for every
configuration file except possibly /etc/master.passwd
For example, /etc/resolv.conf would be turned into a softlink to
/conf/ME/resolv.conf, and /conf/ME itself would be a softlink to
/conf/<HOSTNAME>. The actual resolv.conf configuration file
would reside in /conf/<HOSTNAME>.
If you have a lot of hosts, some configuration files may be commonly
classified. For example, all your shell machines might have the
same /etc/resolv.conf. The solution is to make
/conf/<HOSTNAME>/resolv.conf a softlink to a common directory, say
/conf/HT.SHELL/resolv.conf. It may sound a little messy, but this
sort of categorization actually makes the sysadmins job much, much
easier.
The /conf/ directory hierarchy is stored on the template and
distributed to all the machines along with the rest of the root
partition.
This type of customization is taken from my direct experience
instituting such a system at BEST. At the time, BEST had over 45
machines managed from a single template.
RUN-TIME GENERATED OR MODIFIED FILES IN / or /USR
/etc/aliases.db
/etc/master.passwd
/etc/spwd.db
/etc/pwd.db
/etc/passwd
/etc/namedb/s
/root/.history
/root/.ssh/identity
/root/.ssh/identity.pub
/root/.ssh/random_seed
/root/.ssh/known_hosts
/conf/ME
/kernel* ( note 2 )
/dev ( note 3 )
/var ( note 4 )
/home ( note 4 )
/lost+found
/usr/lost+found
/usr/home ( note 4 )
/usr/crash ( note 5 )
/usr/obj ( note 5 )
/usr/ports ( note 5 )
/usr/src ( note 5 )
/usr/local/crack ( note 5 )
/usr/X11R6/lib/X11/xdm/xdm-errors ( note 6 )
/usr/X11R6/lib/X11/xdm/xdm-pid ( note 6 )
/usr/local/etc/ssh_host_key ( note 6 )
/usr/local/etc/ssh_host_key.pub ( note 6 )
/usr/local/etc/ssh_random_seed ( note 6 )
/conf/ME ( note 7 )
note 2: You typically want to update kernels manually and *NOT*
template them as a safety measure. This also allows you to run
different kernels on different machines or.
note 3: /dev must be updated manually. Some devices, such as tty's and
pty's, use the access and/or modify time and/or user/group
operationally and regenerating the devices on the fly would be
bad.
note 4: /var and /home are usually separately mounted partitions and
thus would not fall under the template, but as a safety measure
the template copier refuse to copy directories named 'home'.
note 5: These are directories that are as often created directly on
/usr as they are separately-mounted partitions. You typically
do not want to template such directories.
note 6: Note that you can solve the problem of xdm and sshd creating
files in /usr. With xdm, edit /usr/X11R6/lib/xdm/xdm-config
and change the errorLogFile and pidFile config lines.
With sshd, add 'HostKey' and 'RandomSeed' directives to specify
/var/db for the location of the host key and run-time sshd
random seed:
HostKey /var/db/ssh_host_key
RandomSeed /var/db/ssh_random_seed
note 7: In this example, /conf/ME is the machine customizer and must
be pointed to the /conf/<full-host-name>/ directory, which is
different for each machine. Thus, the /conf/ME softlink
should never be overwritten by the templating copy.
TYPICAL CUSTOMIZED CONFIGRATION SOFTLINKS
The following files typically need to be turned into softlinks
to /conf/ME/<filename>:
/etc/ccd.conf -> /conf/ME/ccd.conf
/etc/ipfw.conf ...
/etc/fstab
/etc/motd
/etc/resolv.conf
/etc/aliases
/etc/sendmail.cw
/etc/organization
/etc/named.conf
/etc/rc.conf.local
/etc/printcap
/etc/inetd.conf
/etc/login.conf
/etc/gettytab
/etc/ntp.conf
/etc/exports
/root/.k5login -> /conf/ME/root/.k5login
And, of course, /conf/ME is usually a softlink to the appropriate
/conf/<full-host-name>/. Depending on your system configuration,
there may be other files not listed above that you have to worry about.
In many cases, /conf/ME/filename is itself a softlink to
"../HT.xxxx/filename", where HT.xxxx is something like HT.STD ... this
added complexity actually makes it easier to manage multiple
classifications of machines.
DELETION OF FILES
Any file found on the template destination that does not exist in the
source and is not listed as an exception by the source should be deleted.
However, deletion can be dangerous and cpdup will ask for confirmation
by default. Once you know you aren't going to blow things up, you can
turn this feature off and update your systems automatically from cron.
By formalizing the delete operation, you can be 100% sure that it is
possible to recreate / and /usr on any machine with only the original
template and a backup of the ( relatively few ) explicitly-excepted
files. The most common mistake a sysop makes is to make a change to a
file in / or /usr on a target machine instead of the template machine.
If the target machine is updated once a night from cron, the sysop
quickly learns not to do this ( because his changes get overwritten
overnight ). With a manual update, these sorts of mistakes can propogate
for weeks or months before they are caught.
TEMPLATE COPYING AND SAFETY
THE CPDUP PROGRAM
The 'cpdup' program is a program which efficiently duplicates a directory
tree. The program copies source to destination, duplicating devices,
softlinks, hardlinks, files, modification times, uid, gid, flags, perms,
and so forth. The program incorporates several major features:
* The program refuses, absolutely, to cross partition boundries.
i.e. if you were copying the template /usr from an NFS mount to
your /usr, and you had a mount point called /usr/home, the
template copying program would *NOT* descend into /usr/home on
the destination.
This is a safety.
* The program accesses a file called .cpignore in each directory
it descending into on the source to obtain a list of exceptions
for that directory -- that is, files not to copy or mess with.
This is a templating function.
* The program refuses to delete a directory on the destination
being replaced by a softlink or file on the source.
This is a safety mechanism
* The program is capable of maintaing MD5 check cache files and
doing an MD5 check between source and destination during the
scan.
* The program is capable of deleting files/directories on the
destination that do not exist on the source, but asks for
confirmation by default.
This is a templating and a safety mechanism.
* The program uses a copy-to-tmp-and-rename methodology allowing
it to be used to update live filesystems.
This is a templating mechanism.
* The program, by default, tries to determine if a copy is required
by checking modify times, file size, perms, and other stat
elements. If the elements match, it does not bother to copy
( unless an MD5 check is being made, in which case it must read
the destination file ).
You typically run cpdup on the target machine. The target machine
temporarily mounts the template machine's / and /usr via NFS, read-only,
and runs cpdup to update / and /usr. If you use this methodology note
that THERE ARE SECURITY CONSIDERATIONS! See 'SECURITY CONSIDERATIONS WITH
NFS' below.
Whatever script you use that does the NFS mounts should ensure that the
mount succeeded before continuing with the cpdup.
You should create .cpignore files in the appropriate directories on the
template machine's / and /usr partitions so as not to overwrite active
files on the target. The most critical .cpignore files should be
protected with 'chflags schg .cpignore'. Specifically, the ones in /
and /etc, but possibly others as well. For example, the .cpignore
hierarchy for protect /root is:
# /root/.cpignore contains
.history
# /root/.ssh/.cpignore contains
random_seed
known_hosts
authorized_keys
identity
identity.pub
WHEN INITIALLY CONVERTING A TARGET MACHINE TO USE TEMPLATING, ALWAYS
MAKE A FULL BACKUP OF THE TARGET MACHINE FIRST! You may accidently delete
files on the target during the conversion due to forgetting to enter
items into appropriate .cpignore files on the source.
SECURITY CONSIDERATIONS WITH NFS ROOT EXPORT FROM TEMPLATE MACHINE
SECURITY CONSIDERATIONS WITH NFS USR EXPORT FROM TEMPLATE MACHINE
There are some serious security considerations that must be taken into
account when exporting / and /usr on the template machine.
* only export read-only
* the password file ( aka vipw ) may not contain any crypted passwords
at all. You MUST use ssh or kerberos to access the template machine.
You can get away with giving only root a crypted password, but only
if you disallow network root logins and only allow direct root
logins on the console.
* The machine's private ssh_host_key usually resides in /usr/local/etc.
You must move this key to /var/db. You can softlink link so no
modification of sshd_config is required.
* The machine's private ~root/.ssh/identity file is also exposed by
the NFS export, you should move this file to /var/db as well and
put a softlink in ~root/.ssh.
* DON'T EXPORT /var ! Either that, or don't put the private keys
in /var/db ... put them somewhere else.
* You may want to redirect the location of the random_seed file, which
can be done by editing ~root/.ssh/sshd_config and
/usr/local/etc/sshd_config so it is not exposed either.
-Matt
Matthew Dillon
dillon@backplane.com
