On Tue, 2004-12-14 at 09:32 -0600, John Goerzen wrote:
> Is there any practical way I could try to address this?  I would rather
> have the files truncated, or even re-linked to /lost+found or something,
> than have them contain bad data.  I also never seemed to encounter this
> behavior with either ext2 or ext3.  Was I just lucky, or is there
> something fundamentally different about JFS?

I'm not sure how I would address this in jfs.  I don't know about ext3,
but jfs will create and commit a transaction to create a file as soon as
it is opened.  What you want is that no transaction gets committed until
the file is closed.  That would be nearly impossible in jfs since the
transaction creating the file modifies the containing directory, and we
can't hold up pending changes to the directory if other files are being
created before the file is closed.

Mounting with -osync would definitely help, but that would really hurt
performance.  (jfs doesn't fully support the -osync flag, but it will
help in the normal write path.)  Similarly,
tuning /proc/sys/vm/dirty_writeback_centisecs
and /proc/sys/vm/dirty_expire_centisecs should reduce the number of
files affected, but may adversely affect performance as well.

> As for the .so files, I was running apt-get dist-upgrade at the time, so
> they were being created/modified at the time of crash; it was just the
> magnitude of the problem that was startling, especially given that dpkg
> first unpacks things with a temporary filename, then renames them to
> their permanent name to try to avoid any corruption in cases like this.
> 
> I don't know if it does a fsync(), though.

Actually, I think dpkg does call fsync.  I'm not too familiar with
debian, but I wonder whether dpkg unpacks the .so files, or if they get
dynamically created by ld.  I wouldn't expect the files to contain holes
if they were simply unpacked and renamed.  Are the .so files the only
ones with the problem?

Shaggy
-- 
David Kleikamp
IBM Linux Technology Center

_______________________________________________
Jfs-discussion mailing list
[EMAIL PROTECTED]
http://www-124.ibm.com/developerworks/oss/mailman/listinfo/jfs-discussion

Reply via email to