I previously wrote:
> OK, here are the patches described above.
>
> Unfortunately, they haven't been tested. I've given them several
> eyeballings and they appear OK, but when I try to run the ext2 index code
> (even without "-o index" mount option) my system deadlocks somwhere
> inside my ext2i module (tight loop, but SysRQ still works). I doubt
> it is due to these patches, but possibly because I am also working on
> ext3 code in the same kernel. I'm just in the process of getting kdb
> installed into that kernel so I can find out just where it is looping.
I've tested again, now with kdb, and the system loops in ext2_find_entry()
or ext2_add_link(), because there is a directory with a zero rec_len.
While the actual cause of this problem is elsewhere, the fact that
ext2_next_entry() will loop forever with a bad rec_len is a bug not in
the old ext2 code.
I have changed ext2_next_entry() to verify rec_len > EXT2_DIR_REC_LEN(0),
and added a helper function ext2_next_empty() which is nearly the same,
but it follows the de links until it finds one with enough space for a
new entry (used in ext2_add_link()). The former is useful for both Al
and Daniel's code, while the latter may only be useful for Daniel's.
However, the way that I currently fixed ext2_next_entry() is probably
not very safe. If it detects a bad rec_len, we should probably not
touch that block anymore, but it currently just makes it look like the
block is empty. That may lead to deleting the rest of the directory
entries in the block, although this is what e2fsck does anyways... I
at least need to set the error flag in the superblock... Next patch.
I spotted another bug, namely that when allocating a new directory page
it sets rec_len to blocksize, but does not zero the inode field... This
would imply that we would skip a bogus directory entry at the start of
a new page.
As yet I haven't found the cause of the real bug, but at least now my
kernel doesn't lock up on (relatively common) bogus data.
Cheers, Andreas
PS - patch has recieved some editing to remove the other INDEX flag stuff,
so hopefully it is OK
--
Andreas Dilger \ "If a man ate a pound of pasta and a pound of antipasto,
\ would they cancel out, leaving him still hungry?"
http://www-mddsp.enel.ucalgary.ca/People/adilger/ -- Dogbert
This archive was generated by hypermail 2b29 : Tue May 15 2001 - 21:00:25 EST