[linux-lvm] LVM2/DM data corruption during write with 2.6.11.12

[Ludovic Drolez <ldrolez@linbox.com>]

Hi !

We are developing a (GPLed) disk cloning software similar to partimage: it's an
intelligent 'dd' which backups only used sectors. Project info and SVN available 
at http://lrs.linbox.org

Recently I added LVM1/2 support to it, and sometimes we saw LVM restorations
failing randomly (Disk images from RHES are not corrupted, but the result of the
restoration can be lead to a corrupted filesystem). If a restoration fails, just
try another one and it will work...

How the restoration program works:
- I restore the LVM2 administrative data (384 sectors, most of the time),
- I 'vgscan', 'vgchange',
- open for writing the '/dev/dm-xxxx',
- read a compressed file over NFS,
- and put the sectors in place, so it's a succession of '_llseek()' and
'write()' to the DM device.

But, *sometimes*, for example, the current seek position is at 9GB, and some
data is written to sector 0 ! It happens randomly.

Here is a typical strace of a restoration:

write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
_llseek(5, 20963328, [37604032512], SEEK_CUR) = 0
_llseek(5, 0, [37604032512], SEEK_CUR)  = 0
_llseek(5, 2097152, [37606129664], SEEK_CUR) = 0
write(5, "\1\335E\0\f\0\1\2.\0\0\0\2\0\0\0\364\17\2\2..\0\0\0\0\0"..., 512) = 51
2
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
write(5, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 512) = 512
_llseek(5, 88076288, [37694210048], SEEK_CUR) = 0
_llseek(5, 0, [37694210048], SEEK_CUR)  = 0
_llseek(5, 20971520, [37715181568], SEEK_CUR) = 0
write(5, "\377\377\377\377\377\377\377\377\377\377\377\377\377\377"..., 512) = 5
12
....
....

As you can see, there are no seeks to sector 0, but something randomly write
some data to sector 0 !
I could reproduce these random problems on different kind of PCs.

But, the strace above comes from an improved version, which aggregates
'_llseek's. A previous version, which did *many* 512 bytes seeks had much more
problems. Aggregating seeks made the corruption to appears very rarely... And I
more likely to happen, for a 40GB restoration than for a 10GB one.

So less system calls to the DM/LVM2 layer seems to give less corruption probability.


Any ideas ? Newer kernel releases could have fixed such a problem ?


-- 
Ludovic DROLEZ                              Linbox / Free&ALter Soft
www.linbox.com www.linbox.org