hammer2 - update documentation, cleanup

[dragonfly.git] / bin / cpdup / BACKUPS

$DragonFly: src/bin/cpdup/BACKUPS,v 1.4 2007/05/17 08:19:00 swildner Exp $

                            INCREMENTAL BACKUP HOWTO

    This document describes one of several ways to set up a LAN backup and
    an off-site WAN backup system using cpdup's hardlinking capabilities.

    The features described in this document are also encapsulated in scripts
    which can be found in the scripts/ directory.  These scripts can be used
    to automate all backup steps except for the initial preparation of the
    backup and off-site machine's directory topology.  Operation of these
    scripts is described in the last section of this document.


                    PART 1 - PREPARE THE LAN BACKUP BOX

    The easiest way to create a LAN backup box is to NFS mount all your
    backup clients onto the backup box.  It is also possible to use cpdup's
    remote host feature to access your client boxes but that requires root
    access to the client boxes and is not described here.  (But see the
    sections "OFF-SITE BACKUPS" and "SSH SECURITY TIPS" below.)

    Create a directory on the backup machine called /nfs, a subdirectory
    foreach remote client, and subdirectories for each partition on each
    client.  Remember that cpdup does not cross mount points so you will
    need a mount for each partition you wish to backup.  For example:

        [ ON LAN BACKUP BOX ]

        mkdir /nfs
        mkdir /nfs/box1
        mkdir /nfs/box1/home
        mkdir /nfs/box1/var

    Before you actually do the NFS mount, create a dummy file for each
    mount point that can be used by scripts to detect when an NFS mount
    has not been done.  Scripts can thus avoid a common failure scenario
    and not accidently cpdup an empty mount point to the backup partition
    (destroying that day's backup in the process).

        touch /nfs/box1/home/NOT_MOUNTED
        touch /nfs/box1/var/NOT_MOUNTED

    Once the directory structure has been set up, do your NFS mounts and
    also add them to your fstab.  Since you will probably wind up with a
    lot of mounts it is a good idea to use 'ro,bg' (readonly, background
    mount) in the fstab entries.

        mount box1:/home /nfs/box1/home
        mount box1:/var /nfs/box1/var

    You should create a huge /backup partition on your backup machine which
    is capable of holding all your mirrors.  Create a subdirectory called
    /backup/mirrors in your huge backup partition.

        mount <huge_disk> /backup
        mkdir /backup/mirrors


                        PART 2 - DOING A LEVEL 0 BACKUP

    (If you use the supplied scripts, a level 0 backup can be accomplished
    simply by running the 'do_mirror' script with an argument of 0).

    Create a level 0 backup using a standard cpdup with no special arguments
    other then -i0 -s0 (tell it not to ask questions and turn off the
    file-overwrite-with-directory safety feature).  Name the mirror with
    the date in a string-sortable format.

        set date = `date "+%Y%m%d"`
        mkdir /backup/mirrors/box1.${date}
        cpdup -i0 -s0 /nfs/box1/home /backup/mirrors/box1.${date}/home
        cpdup -i0 -s0 /nfs/box1/var /backup/mirrors/box1.${date}/var

    Create a softlink to the most recently completed backup, which is your
    level 0 backup.  Note that using 'ln -sf' will create a link in the
    subdirectory pointed to by the current link, not replace the current
    link. 'ln -shf' can be used to replace the link but is not portable.
    'mv -f' has the same problem.

        sync
        rm -f /backup/mirrors/box1
        ln -s /backup/mirrors/box1.${date} /backup/mirrors/box1

                        PART 3 - DO AN INCREMENTAL BACKUP

    An incremental backup is exactly the same as a level 0 backup EXCEPT
    you use the -H option to specify the location of the most recent 
    completed backup.  We simply maintain the handy softlink pointing at
    the most recent completed backup and the cpdup required to do this
    becomes trivial.

    Each day's incremental backup will reproduce the ENTIRE directory topology
    for the client, but cpdup will hardlink files from the most recent backup
    instead of copying them and this is what saves you all the disk space.

        set date = `date "+%Y%m%d"`
        mkdir /backup/mirrors/box1.${date}
        if ( "`readlink /backup/mirrors/box1`" == "box1.${date}" ) then
            echo "silly boy, an incremental already exists for today"
            exit 1
        endif
        cpdup -H /backup/mirrors/box1 \
              -i0 -s0 /nfs/box1/home /backup/mirrors/box1.${date}/home

    Be sure to update your 'most recent backup' softlink, but only do it
    if the cpdup's for all the partitions for that client have succeeded.
    That way the next incremental backup will be based on the previous one.

        rm -f /backup/mirrors/box1
        ln -s /backup/mirrors/box1.${date} /backup/mirrors/box1

    Since these backups are mirrors, locating a backup is as simple
    as CDing into the appropriate directory.  If your filesystem has a
    hardlink limit and cpdup hits it, cpdup will 'break' the hardlink
    and copy the file instead.  Generally speaking only a few special cases
    will hit the hardlink limit for a filesystem.  For example, the
    CVS/Root file in a checked out cvs repository is often hardlinked, and
    the sheer number of hardlinked 'Root' files multiplied by the number 
    of backups can often hit the filesystem hardlink limit.

                    PART 4 - DO AN INCREMENTAL VERIFIED BACKUP

    Since your incremental backups use hardlinks heavily the actual file
    might exist on the physical /backup disk in only one place even though
    it may be present in dozens of daily mirrors.  To ensure that the
    file being hardlinked does not get corrupted cpdup's -f option can be
    used in conjunction with -H to force cpdup to validate the contents 
    of the file, even if all the stat info looks identical.

        cpdup -f -H /backup/mirrors/box1 ...

    You can create completely redundant (non-hardlinked-dependent) backups
    by doing the equivalent of your level 0, i.e. not using -H.  However I
    do NOT recommend that you do this, or that you do it very often (maybe
    once every 6 months at the most), because each mirror created this way
    will have a distinct copy of all the file data and you will quickly
    run out of space in your /backup partition.

                    MAINTAINANCE OF THE "/backup" DIRECTORY

    Now, clearly you are going to run out of space in /backup if you keep
    doing this, but you may be surprised at just how many daily incrementals
    you can create before you fill up your /backup partition.

    If /backup becomes full, simply start rm -rf'ing older mirror directories
    until enough space is freed up.   You do not have to remove the oldest
    directory first.  In fact, you might want to keep it around and remove
    a day's backup here, a day's backup there, etc, until you free up enough
    space.

                                OFF-SITE BACKUPS

    Making an off-site backup involves similar methodology, but you use
    cpdup's remote host capability to generate the backup.  To avoid
    complications it is usually best to take a mirror already generated on
    your LAN backup box and copy that to the remote box. 

    The remote backup box does not use NFS, so setup is trivial.  Just
    create your super-large /backup partition and mkdir /backup/mirrors.
    Your LAN backup box will need root access via ssh to your remote backup
    box.  See the section "SSH SECURITY TIPS" below.

    You can use the handy softlink to get the latest 'box1.date' mirror 
    directory and since the mirror is all in one partition you can just
    cpdup the entire machine in one command.  Use the same dated directory
    name on the remote box, so:

        # latest will wind up something like 'box1.20060915'
        set latest = `readlink /backup/mirrors/box1`
        cpdup -i0 -s0 /backup/mirrors/$latest remote.box:/backup/mirrors/$latest

    As with your LAN backup, create a softlink on the backup box denoting the
    latest mirror for any given site.

        if ( $status == 0 ) then
            ssh remote.box -n \
                "rm -f /backup/mirrors/box1; ln -s /backup/mirrors/$latest /backup/mirrors/box1"
        endif

    Incremental backups can be accomplished using the same cpdup command,
    but adding the -H option to the latest backup on the remote box.  Note
    that the -H path is relative to the remote box, not the LAN backup box
    you are running the command from.

        set latest = `readlink /backup/mirrors/box1`
        set remotelatest = `ssh remote.box -n "readlink /backup/mirrors/box1"`
        if ( "$latest" == "$remotelatest" ) then
            echo "silly boy, you already made a remote incremental backup today"
            exit 1
        endif
        cpdup -H /backup/mirrors/$remotelatest \
              -i0 -s0 /backup/mirrors/$latest remote.box:/backup/mirrors/$latest
        if ( $status == 0 ) then
            ssh remote.box -n \
                "rm -f /backup/mirrors/box1; ln -s /backup/mirrors/$latest /backup/mirrors/box1"
        endif

    Cleaning out the remote directory works the same as cleaning out the LAN
    backup directory.


                            RESTORING FROM BACKUPS

    Each backup is a full filesystem mirror, and depending on how much space
    you have you should be able to restore it simply by cd'ing into the
    appropriate backup directory and using 'cpdup blah box1:blah' (assuming
    root access), or you can export the backup directory via NFS to your
    client boxes and use cpdup locally on the client to extract the backup.
    Using NFS is probably the most efficient solution.


                        PUTTING IT ALL TOGETHER - SOME SCRIPTS

    Please refer to the scripts in the script/ subdirectory.  These scripts
    are EXAMPLES ONLY.  If you want to use them, put them in your ~root/adm
    directory on your backup box and set up a root crontab.

    First follow the preparation rules in PART 1 above.  The scripts do not
    do this automatically.  Edit the 'params' file that the scripts use
    to set default paths and such.

        ** FOLLOW DIRECTIONS IN PART 1 ABOVE TO SET UP THE LAN BACKUP BOX **

    Copy the scripts to ~/adm.  Do NOT install a crontab yet (but an example
    can be found in scripts/crontab).

    Do a manual lavel 0 LAN BACKUP using the do_mirror script.

        cd ~/adm
        ./do_mirror 0

    Once done you can do incremental backups using './do_mirror 1' to do a
    verified incremental, or './do_mirror 2' to do a stat-optimized
    incremental.  You can enable the cron jobs that run do_mirror and 
    do_cleanup now.

    --

    Setting up an off-site backup box is trivial.  The off-site backup box
    needs to allow root ssh logins from the LAN backup box (at least for
    now, sorry!).  Set up the off-site backup directory, typically
    /backup/mirrors.  Then do a level 0 backup from your LAN backup box
    to the off-site box using the do_remote script.

        cd ~/adm
        ./do_remote 0

    Once done you can do incremental backups using './do_remote 1' to do a
    verified incremental, or './do_mirror 2' to do a stat-optimized
    incremental.  You can enable the cron jobs that run do_remote now.

    NOTE!  It is NOT recommended that you use verified-incremental backups
    over a WAN, as all related data must be copied over the wire every single
    day.  Instead, I recommend sticking with stat-optimized backups
    (./do_mirror 2).

    You will also need to set up a daily cleaning script on the off-site
    backup box.

    SCRIPT TODOS - the ./do_cleanup script is not very smart.  We really
    should do a tower-of-hanoi removal


                              SSH SECURITY TIPS

    To allow root access via ssh, add the following line to your sshd
    configuration on the client boxes (typically /etc/ssh/sshd_config):

        PermitRootLogin forced-commands-only

    If your OpenSSH version is too old to recognize that setting, you
    should update to a more recent version immediately.
    Restart sshd for the settings to take effect.

    On the backup machine, create a special backup key for root:

        mkdir /root/.ssh        # if it doesn't already exist
        cd /root/.ssh
        ssh-keygen -t dsa -N "" -f backup-key

    You now have a key pair, consisting of a secret key called "backup-key"
    and a public key called "backup-key.pub".  The secret key must *NEVER*
    leave the backup machine nor be disclosed in any way!  Note that we
    haven't procted the secret key with a passphrase (-N "") because it
    will be used by cron jobs where no passphrase can be entered.

    On the client boxes, create a file /root/.ssh/authorized_keys.
    It should contain just this line:

        command="/usr/local/bin/cpdup -S",from="<BAKHOST>",no-pty,
        no-port-forwarding,no-X11-forwarding,no-agent-forwarding <PUBKEY>

    This must be on one long line; it has been broken up here for
    readability only.  Note that the options must be separated by commas
    *ONLY* (no spaces).  Replace <BAKHOST> with the IP address or DNS name
    of the backup machine.  Replace <PUBKEY> with the contents of the
    file /root/.ssh/backup-key.pub from the backup machine (the public key,
    not the secret key!).  It typically starts with "ssh-dss" followed by
    a long character sequence that looks like line noise, followed by a
    comment that typically indicates who created the key.

    The format of the authorized_keys file is documented in the sshd(8)
    manual page.  Please refer to it for more details.

    If you have done all of the above correctly, then the root user on the
    backup machine will be able to log into the client boxes as root and
    execute "/usr/local/bin/cpdup -S", but nothing else.

    To further improve security, you can place the slave cpdup on the client
    machine into read-only mode by adding the -R option.  In this case, the
    line from the authorized_keys file should begin as follows:

        command="/usr/local/bin/cpdup -RS",from="<BAKHOST>",etc...

    If you do that, your backup server can only pull backups from the client
    machine, but it won't be able to change anything on it.  That is, you
    cannot use the client machine as a remote target.  So, if an attacker
    manages to be able to execute commands on your backup machine, he won't
    be able to do any harm to your clients.  This also protects against
    human errors, e.g. accidentally swapping source and destination.

    By the way, it doesn't really matter much whether you specify the -R
    option when running cpdup on the backup machine.  If you do it, then
    the -R option will be passed to the slave, but the command="..." entry
    from the authorized_keys file overides it anyway, so the slave always
    runs with the -R option.

    When using cpdup on the backup machine, make sure that the right key is
    used by passing the -i option to the ssh command:

        cpdup -F -i/root/.ssh/backup-key ...

    If one or both of the machines involved has a slow processor, it might
    be worthwhile to use a faster encryption algorithm, for example:

        cpdup -F -cblowfish-cbc ...

    If your OpenSSH version has been patched to support unencrypted transfers
    *AND* you trust the physical network between the machines involved, you
    might want to disable encryption alltogether:

        cpdup -F -cnone ...