Kevin Eilers schrieb:
Hi!
So, it seems that once the mapping is broken, it stays broken until it
is removed and recreated with cryptsetup.
I have experienced something similar. It sometimes took me hours to
produce the corruptions, but once I got the first warnings, every minor
write operation triggered it until I rebooted (root and var filesystems
were affected, so unmounting and remapping didn't work).
Another thing I noticed:
I was using lvm ontop of dm-crypt ontop of raid5 and some filesystems
(logical volumes) would get corrupted when under heavy load, and others
got corrupted when other filesystem were under heavy load. It may be
coincidence but it was always the same combination:
Producing heavy load on the var lv corrupted it.
Producing heavy load on my data lv corrupted the root lv.
All lvs were on the same volume group.
Kevin
Very interesting.
Another observation: I had several occasions where I could stress the
filesystem for hours without a hitch and on the first access after a
pause, the problem would occur immediately. Maybe one could trigger the
problem by the following sequence:
1. cryptsetup
2. mke2fs
3. wait an hour
4. begin copying to the filesystem
And no, the drives were not in a power-save state (which should not be
much of a problem anyway) and I always disable powersaved.
But the most important new info I can contribute is that the Twofish
cipher brings no solution to our problem, so aes-i586 seems fine. Which
is what I expected, since even another DM target (dm-linear) does not
work, which effectively rules out dm-crypt and ciphers completely.
This leaves us with the device mapper (any target) and RAID subsystems
as the main problem candidates. Considering that I had other problems
with cryptoloop, I'd favor the RAID subsystem.
Since I have been using RAID without problems without device mapper or
CRYPTOLOOP, I suspect a race condition (critical region) between those
subsystems.
Hope this helps.
Uwe
---------------------------------------------------------------------
- http://www.saout.de/misc/dm-crypt/
To unsubscribe, e-mail: dm-crypt-unsubscribe@xxxxxxxx
For additional commands, e-mail: dm-crypt-help@xxxxxxxx
