Aug 05, 2018

rsync backups over reverse ssh tunnels

Have been reviewing backups last about 10 years ago (p2, p3), and surprisingly not much changed since then - still same ssh and rsync for secure sync over network remote.

Stable-enough btrfs makes --link-dest somewhat anachronistic, and rsync filters came a long way since then, but everything else is the same, so why not use this opportunity to make things simplier and smoother...

In particular, it was always annoying to me that backups either had to be pulled from some open and accessible port or pushed to same thing on the backup server, which isn't hard to fix with "ssh -R" tunnels - that allows backup server to have locked-down and reasonably secure ssh port open at most, yet provides no remote push access (i.e. no overwriting whatever remote wants to, like simple rsyncd setup would do), and does everything through just a single outgoing ssh connection.

That is, run "rsync --daemon" or localhost, make reverse-tunnel to it when connecting to backup host and let it pull from that.

On top of being no-brainer to implement and use - as simple as ssh from behind however many NATs - it avoids following (partly mentioned) problematic things:

  • Pushing stuff to backup-host, which can be exploited to delete stuff.
  • Using insecure network channels and/or rsync auth - ssh only.
  • Having any kind of insecure auth or port open on backup-host (e.g. rsyncd) - ssh only.
  • Requiring backed-up machine to be accessible on the net for backup-pulls - can be behind any amount of NAT layers, and only needs one outgoing ssh connection.
  • Specifying/handling backup parameters (beyond --filter lists), rotation and cleanup on the backed-up machine - backup-host will handle all that in a known-good and uniform manner.
  • Running rsyncd or such with unrestricted fs access "for backups" - only runs it on localhost port with one-time auth for ssh connection lifetime, restricted to specified read-only path, with local filter rules on top.
  • Needing anything beyond basic ssh/rsync/python on either side.

Actual implementation I've ended up with is ssh-r-sync + ssh-r-sync-recv scripts in fgtk repo, both being rather simple py3 wrappers for ssh/rsync stuff.

Both can be used by regular uid's, and can use rsync binary with capabilities or sudo wrapper to get 1-to-1 backup with all permissions instead of --fake-super (though note that giving root-rsync access to uid is pretty much same as "NOPASSWD: ALL" sudo).

One relatively recent realization (coming from acme-cert-tool) compared to scripts I wrote earlier, is that using bunch of script hooks all over the place is a way easier than hardcoding a dozen of ad-hoc options.

I.e. have option group like this (-h/--help output from argparse):

Hook options:
  -x hook:path, --hook hook:path
    Hook-script to run at the specified point.
    Specified path must be executable (chmod +x ...),
      will be run synchronously, and must exit with 0
      for tool to continue operation, non-0 to abort.
    Hooks are run with same uid/gid
      and env as the main script, can use PATH-lookup.
    See --hook-list output to get full list of
      all supported hook-points and arguments passed to them.
    Example spec: -x rsync.pre:~/hook.pre-sync.sh
  --hook-timeout seconds
    Timeout for waiting for hook-script to finish running,
      before aborting the operation (treated as hook error).
    Zero or negative value will disable timeout. Default: no-limit
  --hook-list
    Print the list of all supported
      hooks with descriptions/parameters and exit.

And --hook-list providing full attached info like:

Available hook points:

  script.start:
    Before starting handshake with authenticated remote.
    args: backup root dir.

  ...

  rsync.pre:
    Right before backup-rsync is started, if it will be run.
    args: backup root dir, backup dir, remote name, privileged sync (0 or 1).
    stdout: any additional \0-separated args to pass to rsync.
      These must be terminated by \0, if passed,
        and can start with \0 to avoid passing any default options.

  rsync.done:
    Right after backup-rsync is finished, e.g. to check/process its output.
    args: backup root dir, backup dir, remote name, rsync exit code.
    stdin: interleaved stdout/stderr from rsync.
    stdout: optional replacement for rsync return code, int if non-empty.

Hooks are run synchronously,
  waiting for subprocess to exit before continuing.
All hooks must exit with status 0 to continue operation.
Some hooks get passed arguments, as mentioned in hook descriptions.
Setting --hook-timeout (defaults to no limit)
  can be used to abort when hook-scripts hang.

Very trivial to implement and then allows to hook much simplier single-purpose bash scripts handling specific stuff like passing extra options on per-host basis, handling backup rotation/cleanup and --link-dest, creating "backup-done-successfully" mark and manifest files, or whatever else, without needing to add all these corner-cases into the main script.

One boilerplate thing that looks useful to hardcode though is a "nice ionice ..." wrapper, which is pretty much inevitable for background backup scripts (though cgroup limits can also be a good idea), and fairly easy to do in python, with minor a caveat of a hardcoded ioprio_set syscall number, but these pretty much never change on linux.

As a side-note, can recommend btrbk as a very nice tool for managing backups stored on btrfs, even if for just rotating/removing snapshots in an easy and sane "keep A daily ones, B weekly, C monthly, ..." manner.

[code link: ssh-r-sync + ssh-r-sync-recv scripts]

Member of The Internet Defense League