Andrew Tridgell and Stephen C.  Tweedie have reported two different Oopses
caused by a race condition in the mbcache, which is responsible for
extended attribute sharing in ext2 and ext3.  Stephen tracked down the bug;
I did the fix.

Explanation:

The mbcache caches the locations and content hashes of xattr blocks.  There
are two access strategies: [1] xattr block disposal via
mb_cache_entry_get(), [2] xattr block reuse (sharing) via
mb_cache_entry_find_{first,next}().  There is no locking between the two
methods, so between one mb_cache_entry_find_x and the next, a
mb_cache_entry_get might come in, unhash the cache entry, and change the
journaling state of the xattr buffer.  Subsequently, two things can happen:
[a] the next mb_cache_entry_find_x may try to follow the mbcache hash chain
starting from the entry that has become unhashed, which now is a stale
pointer, [b] the block may have become deallocated, and then we try to
reuse it.

Fix this by converting the mbcache into a readers-writer style lock, and
protect all block accesses in ext2/ext3 by the mbcache entry lock.  This
ensures that destroying blocks is an exclusive operation that may not
overlap xattr block reuse, while allowing multiple "re-users".  Write
access to the xattr block's buffer is protected by the buffer lock.  
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>