Linus,
    Submitted this to Al a few days ago with no response.  I later ran 
across a message from him saying that this was on his list of things to 
do, so I hope I'm not duplicating work.

The following patch moves the BKL out of notify_change() and into the
individual filesystems.  There was an attempt to remove the BKL from the
function here:
http://groups.google.com/groups?hl=en&selm=20020118043308.B9A75B581%40smtp.transgeek.com.lucky.linux.kernel
However, it looks to me like Craig's I_ATTR_LOCK bit in i_state plus the
inode_lock just about equals a semaphore.  Craig agreed that a semaphore 
  was equivalent, but pointed out that his approach is more efficient 
with space than adding a semaphore to struct inode.

My patch adds a semaphore to the inode structure, removes the BKL from
notify_change(), and shifts the BKL into the individual filesystems'
setattr() functions.  I doubt that many of them actually need it, but I
figure the FS maintainers can remove it.  When I posted the original 
version, Richard Gooch let me know that devfs didn't need it, so I 
removed that part.

The BKL use in notify_change() is replaced by the semaphore i_attr_lock. 
  It might be possible to use i_sem instead, but at least one filesystem 
(UMSDOS) already uses i_sem in its notify function.

-- 
Dave Hansen
haveblue@us.ibm.com