On Sun, 2007-07-15 at 12:59 -0700, Andrew Morton wrote:

On Sun, 15 Jul 2007 21:21:03 +0200 Peter Zijlstra [off-list ref] wrote:

quoted

Shows the current stacktrace where we violate the previously established
locking order.

yup, but the lock_page() which we did inside truncate_mutex was a 
lock_page() against a different address_space: the blockdev mapping.

So this is OK - we'll never take truncate_mutex against the blockdev
mapping (it doesn't have one, for a start ;))

This is similar to the quite common case where we take inode A's
i_mutex inside inode B's i_mutex, which needs special lockdep annotations.

I think.  I haven't looked into this in detail.

Right, I can make lock_page classes per address space. Lets see if this
one goes away.