bin/121684: : dump(8) frequently hangs

[Jeremy Chadwick]

On Mon, Sep 01, 2008 at 02:38:56PM -0700, Jeremy Chadwick wrote:
> ...  I'll try to find the thread (it was a year or so ago)
> where a developer told me more or less what was going on.  The problem
> was that UFS2 snapshot generation, over time, becomes slower and slower
> to generate (this is what dump does on UFS2 systems, with or without the
> -L flag), and is a known design issue.

The issue I'm describing above is fully discussed here.  Be sure to read
the threads I reference in my mail, as well as Eric Anderson and Peter
Jeremy's replies:

http://lists.freebsd.org/pipermail/freebsd-hackers/2007-October/021985.html

There's also a statement from Kris Kennaway in one of the above
referenced threads, who states that "UFS2 snapshots (e.g. dump -L) were
not intended to be used this way":

http://lists.freebsd.org/pipermail/freebsd-stable/2007-January/032070.html

In that same thread, Doug Ambrisko provided a patch which supposedly
made UFS2 snapshot generation significantly better for him, specifically
with regards to dump "getting wedged":

http://lists.freebsd.org/pipermail/freebsd-stable/2007-January/032250.html

Another thread discussing this problem ("lock up when dump is used"):

http://lists.freebsd.org/pipermail/freebsd-stable/2008-June/042823.html

And finally, an issue I encountered (rebooting a system in the middle of
dump -L) was documented here:

http://lists.freebsd.org/pipermail/freebsd-stable/2007-September/036916.html

I'll be adding these issues to my FreeBSD wiki page momentarily.

I think folks should reconsider use of dump when UFS2 filesystems are
chosen.  This goes against what the O'Reilly "Backup & Recovery" book
concludes, and what Elizabeth Zwicky concluded in her backup/recovery
torture tests (albeit 17 years ago, but the document is referenced in
the FreeBSD handbook itself):
http://www.coredumps.de/doc/dump/zwicky/testdump.doc.html

If you're using older UFS1 filesystems, dump should be fine, as there's
no snapshot generation supported there.

If the problem is truly with UFS2 snapshot generation, then I'd like to
know why dump bothers with spitting out a warning when -L is used, more
or less inducing people to modify their scripts to use the -L flag.  The
commit which induced this behaviour is from November 2002, and modified
in April 2004 to advocate -L usage even more:

http://www.freebsd.org/cgi/cvsweb.cgi/src/sbin/dump/main.c.diff?r1=1.42;r2=1.43;f=h
http://www.freebsd.org/cgi/cvsweb.cgi/src/sbin/dump/main.c.diff?r1=1.57;r2=1.58;f=h

-- 
| Jeremy Chadwick                                jdc at parodius.com |
| Parodius Networking                       http://www.parodius.com/ |
| UNIX Systems Administrator                  Mountain View, CA, USA |
| Making life hard for others since 1977.              PGP: 4BD6C0CB |