list_move(&inode->i_list, &inode_in_use);
 		} else {
 			list_move(&inode->i_list, &inode_unused);
+			inodes_stat.nr_unused++;
 		}
 	}
 	wake_up_inode(inode);

Are you sure the above diff is correct? It was added somewhere between
2.6.5 and 2.6.8. I think it's wrong.

The only way I can imagine the i_count to be zero in the above path, is
that I_WILL_FREE is set.  And if I_WILL_FREE is set, then we must not
increase nr_unused.  So I believe the above change is buggy and it will
definitely overstate the number of unused inodes and it should be backed
out.

Note that __writeback_single_inode before calling __sync_single_inode, can
drop the spinlock and we can have both the dirty and locked bitflags clear
here:

		spin_unlock(&inode_lock);
		__wait_on_inode(inode);
		iput(inode);
XXXXXXX
		spin_lock(&inode_lock);
	}
	use inode again here

a construct like the above makes zero sense from a reference counting
standpoint.

Either we don't ever use the inode again after the iput, or the
inode_lock should be taken _before_ executing the iput (i.e. a __iput
would be required). Taking the inode_lock after iput means the iget was
useless if we keep using the inode after the iput.

So the only chance the 2.6 was safe to call __writeback_single_inode
with the i_count == 0, is that I_WILL_FREE is set (I_WILL_FREE will
prevent the VM to free the inode in XXXXX).

Potentially calling the above iput with I_WILL_FREE was also wrong
because it would recurse in iput_final (the second mainline bug).

The below (untested) patch fixes the nr_unused accounting, avoids recursing
in iput when I_WILL_FREE is set and makes sure (with the BUG_ON) that we
don't corrupt memory and that all holders that don't set I_WILL_FREE, keeps
a reference on the inode!

Signed-off-by: Andrea Arcangeli <andrea@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>