Again: Inconsistent snapshot contents
Titus von Boxberg
ut at bhi-hamburg.de
Wed Jun 8 14:11:34 GMT 2005
I sent a mail about inconsistent snapshot contents a week ago but did not
get a response yet.
Please notify me at least if this is the wrong mailing list for these
This is the abstract:
On two servers (5.2.1-RELEASE and 5.3-RELEASE) we use the following
mksnap_ffs / /.snap/backupsnap
mdconfig -a -t vnode -f /.snap/backupsnap -u 5
mount -o ro /dev/md5 /mnt
cp fileinquestion /tmp/x1 # see below
tar cf /dev/sa0 ....
cp fileinquestion /tmp/x2 # see below
tar df /dev/sa0
if [ $? -eq 1 ] ; then # see below
tar -C /tmp -xf /dev/sa0 fileinquestion
mv /tmp/fileinquestion /tmp/x3
I cannot see anything wrong with the script or the underlying
assumptions (snapshots are COMPLETE consistent views of the state of
the file system at a certain instant).
I think it's highly unlikely that the tar command(s) are able
to alter the read only mounted snapshot.
The verify tar almost always complains that at least one
(and always the same) file(s) is/are differing.
I tracked the file contents of one file that the verify
tar usually complains about and could see that its contents
ON THE SNAPSHOT (but NOT its size)
frequently differ over time.
The contents at least of this file in question are
modified every minute or so OUTSIDE (of course) the snapshot.
It does not have anything to do with the tape. The problem
also occurs when tarring onto or from a disk file on a holding disk.
I'm also quite sure that this problem is NOT reproducible
if the backup process (tar) is not run.
It is not sufficient to create the snapshot and compare
<fileinquestion> over time ***without also starting
the backup command***.
The hardware configs of both servers are quite different
(AthlonXP2800/512/ATA60GB vs. Celeron600/256/ATARAID 80GB).
CONCLUSION 1 -->
I'm quite sure now that the contents of the snapshot may be
inconsistent over time.
CONCLUSION 2 -->
There is some bug in the kernel maybe related to a side effect
that occurs somehow when the tar commands are run
AND snapshotted files change (outside the snapshot)
in a certain manner (by content but apparently NOT in size).
Is something wrong with the ASSUMPTIONS above?
Have I overlooked some bugfix between 5.3-RELEASE and 5-STABLE?
Is this the wrong mailing list?
Should I post this question / problem to some other list?
My FreeBSD kernel specific skills are currently restricted to
writing device drivers, but I'm willing and able to help by
coding and installing or using any hook or debugging routine
into the ufs kernel code if someone could point me to the right place.
This would be the optimum answer on this mail.
Thanks in advance.
More information about the freebsd-fs