Bug in extents+mballoc+delalloc

Hello

I am using delalloc+mballoc+extent (from ftp://ftp.clusterfs.com/pub/people/alex/2.6.16.8)
with kernel 2.6.16.20 and found that some of my data gets corrupted.
Unfortunately I am unable to reproduce this reliably. The symptoms are
that files which lie for a longer time in a directory are suddenly
corrupted after they held the content for several hours correctly.
In one case the ~/.ssh/known_hosts file was always loosing host keys
that where added. It always lost them after approx. 12 hours. Copying
the ~/.ssh/known_hosts to ~/.ssh/known_hosts.tmp and then back again
solved the problem. Similar errors occur with the ~/.bash_history,
when I log in the next day all the history from the previous day is gone.

One of my application uses mmap() to create a file and stores a linked list
in there. When an element get added or removed the length of the file is
adapted. At the beginning there is a int which tells me how many elements
are in there. Often this number of elements is incorrect when the
number of elements is reduced. Adding an element is not a problem.
However I sometimes saw that the length of this file is way to short
when doing an 'ls -al' or a stat() on it. But when other programs
map to it they see the full content. What is funny, another of my
application is doing the same thing with a structure not a linked
list as elements. And here I have so far not discovered any problems.
The linked list I do however always align to a int.

This sounds all very confusing and I did try to create a reproducible
test case but failed. So what can I do to help find the bug?

The system is a four CPU Opteron running a 64bit Fedora Core 5. Under
the filesystem there is a software raid 1+0 over eight disks. The system
is used to distribute files (2.7 million files with 627 GB daily).
None of the users have so far complained about corrupted files.
Only files that lie around for a longer time seem to be affected.

Please CC me since I am not assigned to this list.

Thanks,
Holger
-- 


Using Tomcat but need to do more? Need to support web services, security?
Get stuff done quickly with pre-integrated technology to make your job easier
Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642