[PATCH] iommu fixes, round 3

[Benjamin Herrenschmidt]

> :) Respectfully, I still disagree.  The caller is a procfs-specific function
> related to an interface that we're hoping to deprecate soon.  We want this to
> happen any time a node is removed, not anytime a node is removed using
> interface so-and-so.
> 
> To me, it makes sense to put this here since of_add_node() calls 
> of_finish_node_dynamic(), which creates the table.

I hate that interface... but I suppose we can merge the patch for now. I
think this should be changed tho. It's no business of the low level
device-tree manipulation functions to know about such things as iommu
tables. And what will happen the day I remove the iommu table pointer
from the struct device-node anyway ?

If your interface to userland relies on that, then it's broken and will
have to be reworked :( Maybe we can get away be creating a notifier
mecanism for something in the kernel to get called back after nodes are
beeing added and before they are beeing removed, that would be ok I
suppose, but the low level tree manipulation has to stay separate. I do
intend, in the long run, to remove all those additional fields we put in
struct device-tree...

Ben.